Generic season index - no season is length 0 (#1845)

<!--Please ensure the PR fulfills the following requirements! -->
<!-- If this is your first PR, make sure to add your details to the
AUTHORS.rst! -->
### Pull Request Checklist:
- [ ] This PR addresses an already opened issue (for bug fixes /
features)
    - This PR fixes #xyz
- [x] Tests for the changes have been added (for bug fixes / features)
- [x] (If applicable) Documentation has been added / updated (for bug
fixes / features)
- [x] CHANGELOG.rst has been updated (with summary of main changes)
- [x] Link to issue (:issue:`number`) and pull request (:pull:`number`)
has been added

### What kind of change does this PR introduce?

* Implements `generic.season`, a generic index function for computing
the start/end/length of seasons. Based on `run_length.season` but with
units management and the comparison.
* Divide `run_length.season` into `season_start`, `season_end`, plus the
existing `season_length`. The idea of grouping code together was good,
but it added a lot of unnecessary computations when only one facet is
needed (i.e. most cases). I think I was able to divide it elegantly,
using only one magic fast path.
* Rewrite `frost_free_season_start` and `frost_free_season_end`. Those
were not respecting the "season" definition. Now they do.

### Does this PR introduce a breaking change?
`frost_free_season_start` and `frost_free_season_end` have changed.
Their notion of "season" is not more strict. I think the difference not
significative for most cases. No tests needed a change, but again our
tests are very simple.

All season length indicators will now return 0 when no season is found
(no start). Previously, when no start was found we returned 0 when an
end was found and `nan` otherwise. I think the idea was to assume that
neither end nor start meant the data was probably wrong. But that's the
job of the missing checks, not the `run_length` helper. Thus I applied
the more logical rule : no season == 0 length.

I have reverted some changes of #1796, I'm sorry. I'm trying to
generalize the indicators and it made more sense that all aspects of the
same "growing season" were using the same exact arguments.

### Other information:
This allows indices to use run length function that return a DataArray
after being called with a DataArray. This "type preservation" will be
quite handy in my next PR, stay tuned!

Sadly, this branch is based off #1838 because I need both for PC.


### Ugh
The amount of copied docstring text and signatures is ridiculous. I
really want to remove all those one-liner indices and only keep the
indicators. But I fear this would be too breaking of a change as it
would remove elements from the API.
Right now, we have inconsistent indicators with inconsistent
documentation because everything was copy-pasted by hand with
incremental changes that were not always ported back to the other
indices.

CHANGELOG.rst

@@ -16,6 +16,11 @@ Breaking changes
 ^^^^^^^^^^^^^^^^
 * As updated in ``cf_xarray>=0.9.3``, dimensionless quantities now use the "1" units attribute as specified by the CF conventions, previously an empty string was returned. (:pull:`1814`).
 
+Breaking changes
+^^^^^^^^^^^^^^^^
+* The definitions of the ``frost_free_season_start`` and ``frost_free_season_end`` have been slightly changed to be coherent with the ``frost_free_season_length`` and `xclim`'s notion of `season` in general. Indicator and indices signature have changed. (:pull:`1845`).
+* Season length indicators have been modified to return ``0`` for all cases where a proper season was not found, but the data is valid. Previously, a ``nan`` was given if neither a start or an end were found, even if the data was valid, and a ``0`` was given if an end was found but no start. (:pull:`1845`).
+
 Bug fixes
 ^^^^^^^^^
 * Fixed the indexer bug in the ``xclim.indices.standardized_index_fit_params`` when multiple or non-array indexers are specified and fitted parameters are reloaded from netCDF. (:issue:`1842`, :pull:`1843`).
@@ -26,6 +31,7 @@ Internal changes
 ^^^^^^^^^^^^^^^^
 * Changed the French translation of "wet days" from "jours mouillés" to "jours pluvieux". (:issue:`1825`, :pull:`1826`).
 * In order to adapt to changes in `pytest`, the doctest fixtures have been split from the main testing suite and doctests are now run using ``$ python -c 'from xclim.testing.utils import run_doctests; run_doctests()'``. (:pull:`1632`).
+* Added ``xclim.indices.generic.season`` to make season start, end, and length indices. Added a ``stat`` argument to ``xclim.indices.run_length.season`` to avoid returning a dataset. (:pull:`1845`).
 
 CI changes
 ^^^^^^^^^^

tests/test_indices.py

@@ -1160,7 +1160,7 @@ class TestGrowingSeasonLength:
         [
             ("1950-01-01", "1951-01-01", 0),  # No growing season
             ("2000-01-01", "2000-12-31", 365),  # All year growing season
-            ("2000-07-10", "2001-01-01", np.nan),  # End happens before start
+            ("2000-07-10", "2001-01-01", 0),  # End happens before start
             ("2000-06-15", "2001-01-01", 199),  # No end
             ("2000-06-15", "2000-07-15", 31),  # Normal case
         ],

tests/test_run_length.py

@@ -470,7 +470,7 @@ class TestRunsWithDates:
         [
             ("07-01", 210, 70),
             ("07-01", 190, 50),
-            ("04-01", 150, np.nan),  # date falls early
+            ("04-01", 150, 0),  # date falls early
             ("11-01", 150, 165),  # date ends late
             (None, 150, 10),  # no date, real length
         ],
@@ -492,7 +492,7 @@ def test_season_length(self, tas_series, date, end, expected, use_dask, ufunc):
             runs,
             window=1,
             dim="time",
-            date=date,
+            mid_date=date,
         )
         np.testing.assert_array_equal(np.mean(out.load()), expected)
 
@@ -625,7 +625,7 @@ def test_run_with_dates_different_calendars(self, calendar, expected):
         out = (
             (tas > 0)
             .resample(time="YS-MAR")
-            .map(rl.season_length, date="03-02", window=2)
+            .map(rl.season_length, mid_date="03-02", window=2)
         )
         np.testing.assert_array_equal(out.values[1:], [250, 250])
 
@@ -643,9 +643,7 @@ def test_run_with_dates_different_calendars(self, calendar, expected):
         )
         np.testing.assert_array_equal(out.values[1:], np.array(expected) + 1)
 
-    @pytest.mark.parametrize(
-        "func", [rl.first_run_after_date, rl.season_length, rl.run_end_after_date]
-    )
+    @pytest.mark.parametrize("func", [rl.first_run_after_date, rl.run_end_after_date])
     def test_too_many_dates(self, func, tas_series):
         tas = tas_series(np.zeros(730), start="2000-01-01")
         with pytest.raises(ValueError, match="More than 1 instance of date"):

tests/test_temperature.py

@@ -368,7 +368,7 @@ def test_real_data(self, atmosds):
 
 class TestFrostSeasonLength:
     def test_simple(self, tasmin_series):
-        a = np.zeros(730) + K2C + 15
+        a = np.zeros(731) + K2C + 15
         a[300:400] = K2C - 5
         a[404:407] = K2C - 5
         tasmin = tasmin_series(a, start="2000-01-01")
@@ -380,7 +380,7 @@ def test_simple(self, tasmin_series):
         np.testing.assert_array_equal(out, [np.nan, 100, np.nan])
 
         out = atmos.frost_season_length(tasmin=tasmin, mid_date="07-01", freq="YS")
-        np.testing.assert_array_equal(out, [np.nan, np.nan])
+        np.testing.assert_array_equal(out, [0, 181])
 
 
 class TestColdSpellDays: