Respect average nthreads in adaptive

[mrocklin]

Previously we would assume in adaptive that workers had a single core.

Now we do a pass through all the workers to sum up the number of threads per machine and divide the recommendation by this amount.

This is not very expensive

In [1]: class Worker:
   ...:     def __init__(self):
   ...:         self.cpu = 4
   ...:

In [2]: workers = [Worker() for _ in range(1000)]

In [3]: %timeit sum(w.cpu for w in workers)
33.6 µs ± 170 ns per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

Supercedes #8039

[github-actions]

Unit Test Results

See test report for an extended history of previous test failures. This is useful for diagnosing flaky tests.

       20 files  ±  0         20 suites  ±0   10h 37m 14s ⏱️ - 46m 54s
  3 745 tests +  2    3 633 ✔️ +  2     106 💤 ±  0    6 ❌ ±0 
36 135 runs   - 69  34 383 ✔️  - 57  1 738 💤  - 10  14 ❌  - 2 

For more details on these failures, see this check.

Results for commit 2a2aedf. ± Comparison against base commit 9d9702e.

This pull request removes 1 and adds 3 tests. Note that renamed tests count towards both.

distributed.deploy.tests.test_adaptive ‑ test_target_duration

distributed.deploy.tests.test_adaptive ‑ test_respect_average_nthreads
distributed.deploy.tests.test_adaptive ‑ test_target_duration[1]
distributed.deploy.tests.test_adaptive ‑ test_target_duration[5]

♻️ This comment has been updated with latest results.

[mrocklin]

@fjetter handing this off to you and your team if you're cool with that.

[fjetter]

@hendrikmakait feel free to merge once the test works

[mrocklin]

Thank you both for getting this in.

[mrocklin]

cc @dchudz , this is in.