fix(python): DataFrame rows_by_key returning key tuples with elements in wrong order

[lukemanley]

Fixes the following and simplifies the implementation a bit.

In [1]: import polars as pl

In [2]: df = pl.DataFrame(
   ...:    ...:     {
   ...:    ...:         "a": ["a", "a"],
   ...:    ...:         "b": ["b", "b"],
   ...:    ...:         "c": [1, 2],
   ...:    ...:     }
   ...:    ...: )

In [3]: df.rows_by_key(["a", "b"])
Out[3]: defaultdict(list, {('a', 'b'): [1, 2]})   # <-- key tuple ordered as (a, b) as requested

In [4]: df.rows_by_key(["b", "a"])
Out[4]: defaultdict(list, {('a', 'b'): [1, 2]})   # <-- key tuple ordered as (a, b) rather than (b, a) as requested

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 79.55%. Comparing base (5f11dd9) to head (c378400).
Report is 32 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff             @@
##             main   #19486       +/-   ##
===========================================
+ Coverage   59.87%   79.55%   +19.68%     
===========================================
  Files        1545     1545               
  Lines      213433   213405       -28     
  Branches     2442     2429       -13     
===========================================
+ Hits       127791   169782    +41991     
+ Misses      85092    43075    -42017     
+ Partials      550      548        -2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[alexander-beedie]

Thanks; I'll take a look at this one later tonight (or tomorrow).
Looks much simplified indeed 👍

[lukemanley]

@alexander-beedie - sorry for the ping, just checking in on this.

[alexander-beedie]

Apologies, got distracted and let this slip through the cracks - taking a look now.

[alexander-beedie]

Tested various parameter permutations on a large-scale frame, and 5 of 7 combinations were faster with this new approach. Really like the simplification you made here.

I might take a look later to see if we can get the two permutations that were marginally slower (~15%) back to the same speed, but as the other 5 parameter combinations were about the same amount faster, this PR looks like a clear win to me (in addition to the correctness fix). Great job 👌