Restore dict based typesafe caching

[ethanhs]

This also restores isfile_case caching and fixes a type error (and legitimate bug) (!).

The last remaining follow up from #4693 is to run only one Python subprocess.

[gvanrossum]

Does this need to go into release-0.590? I.e. is it a critical bugfix to the PEP 561 code or just a performance improvement? (I haven't looked yet -- was there a perf regression in situations that don't invoke PEP 561?)

[ethanhs]

@gvanrossum this change (just this!) is a bug on master. lib_path should be a tuple, and as you can see to the left, it is passed as a list. I can split this fix out if you don't want the rest. (This could cause issues with lru_cache).

(The rest is related to file system consistency)

EDIT: added more info.

[JelleZijlstra]

What is the concrete effect here? Could this cause problems to users?

[ethanhs]

Essentially, lru_cache won't be able to hash the passed list.

$ python -m mypy.stubgen --no-import mypy
Traceback (most recent call last):
  File "C:\Python36\lib\runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "C:\Python36\lib\runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "C:\Users\ethanhs\Documents\mypy\mypy\stubgen.py", line 1022, in <module>
    main()
  File "C:\Users\ethanhs\Documents\mypy\mypy\stubgen.py", line 901, in main
    raise e
  File "C:\Users\ethanhs\Documents\mypy\mypy\stubgen.py", line 898, in main
    include_private=options.include_private)
  File "C:\Users\ethanhs\Documents\mypy\mypy\stubgen.py", line 106, in generate_stub_for_module
    interpreter=interpreter)
  File "C:\Users\ethanhs\Documents\mypy\mypy\stubgen.py", line 164, in find_module_path_and_all
    interpreter)
TypeError: unhashable type: 'list'

So yes, at least changing the list to a tuple should go into 0.590.

[JelleZijlstra]

Then maybe make a separate PR for that change?

[gvanrossum]

OK, please split that out -- I can cherry-pick a 3-line fix to stubgen with a lot less risk and effort than this whole PR.

[ethanhs]

@gvanrossum #4899

[JukkaL]

LGTM! Thanks for restoring the old caching implementation -- when most of the code can be type checked, gaps in type checking coverage are dangerous as people tend to assume that type checking will catch certain categories of errors.