[dask] reduce test times

[jameslamb]

This PR proposes a refactoring of tests/python_package_test/test_dask.py, to reduce the runtime of the Dask tests.

• makes models smaller by setting n_estimators=10 and num_leaves=10 (the defaults are n_estimators=100 and num_leaves=31)
• adds the predict_proba() tests for DaskLGBMClassifier into the existing test_classifier test. I don't think effectively testing predict() and predict_proba() requires separate training runs

Impact of this change

I tried running pytest tests/python_package_test/test_dask.py 5 times on master and 5 times on this branch (on my laptop). I found that the changes in this PR reduce the runtime of the Dask tests by around 35 seconds.

I think this speedup will have a meaningful impact on development speed for the Dask module.

before (avg = 137s):

============================ 29 passed, 2 warnings in 150.19s (0:02:30) =============================
============================= 29 passed, 1 warning in 140.21s (0:02:20) =============================
============================ 29 passed, 2 warnings in 132.08s (0:02:12) =============================
============================ 29 passed, 2 warnings in 132.68s (0:02:12) =============================
============================ 29 passed, 2 warnings in 133.04s (0:02:13) =============================

after combining plus less iterations (avg = 102s):

============================= 23 passed, 1 warning in 101.96s (0:01:41) =============================
============================= 23 passed, 1 warning in 103.60s (0:01:43) =============================
============================= 23 passed, 1 warning in 101.74s (0:01:41) =============================
============================= 23 passed, 1 warning in 101.34s (0:01:41) =============================
============================= 23 passed, 1 warning in 102.70s (0:01:42) =============================

Notes for Reviewers

1. I intentionally did not change n_estimators in the test_regression test. That test requires a lot of iterations from distributed and local training to look similar, and I found that they produced very different results for small values of n_estimators.
2. If this PR is accepted, I'll add a review comment with similar suggestions on [python-package] [dask] Add DaskLGBMRanker #3708

[StrikerRUS]

LGTM! Thanks! I find this useful given the growing number of CI jobs and services.

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.