[python-package] [ci] remove runtime docstring building, test with different PYTHONOPTIMIZE levels #7403

Fixes #7362
Contributes to #7327

The Python package has 7 scikit-learn compatible estimators, and there is a lot of overlap in their docstrings. To keep them consistent, the estimators reference build up their docstrings at import time from templates and inherited docstrings, like this:

https://github.com/lightgbm-org/LightGBM/blob/3cd99268f09615b2e2da6e93269a7e6db199b342/python-package/lightgbm/sklearn.py#L1555-L1561

This proposes removing that stuff in favor of just using literal docstrings.

Changes in this PR

• removes all the import-time docstring-templating stuff, in favor of literal docstrings


• adds unit tests to ensure the docstrings remain consistent with each other


• adds tests in CI that import lightgbm works when run under PYTHONOPTIMIZE=1 and PYTHONOPTIMIZE=2
  

Benefits of this change

• allows lightgbm to be imported when using PYTHONOPTIMIZE=2 (fixes #7362)


• very slightly reduces import time (in my local testing, by around 10ms)
  


• makes it easier to reason about the docs and edit them

How I tested this

• all those unit tests I added


• import-timing code from #7327


• built the docs on readtehdocs
  
  
  • build: https://app.readthedocs.org/projects/lightgbm/builds/?version__slug=python-no-runtime-doc-parsing
    
  
  
  • docs: https://lightgbm.readthedocs.io/en/python-no-runtime-doc-parsing