Fine-grained imports management

[layus]

Currently we only care about toplevel imports, but some packages provide different subparts of a namespace (optional dependencies/extras), or share the same toplevel namespace (e.g. google...).

To do this we may need to rewrite our own version of isort.

[layus]

Might also fix #419 if we get rid of isort

[layus]

Searching for import scanning libraries, I found:

(rust) https://github.com/saadmk11/python-third-party-imports
(python) https://github.com/bazel-contrib/rules_python/blob/main/gazelle/modules_mapping/generator.py
(mypy) https://mypy.readthedocs.io/en/stable/running_mypy.html#how-imports-are-found
(python) https://pypi.org/project/py-module-parser/, maybe a bit too much
(native python parser) https://stackoverflow.com/questions/44988487/regex-to-parse-import-statements-in-python#answer-44988778

I am really interested in the native python parser. I fail to see what can be so hard about sorting imports that we could not reimplement it ourselves. I have the feeling that isort is overkilling it.

[layus]

Also, the reference logic: https://github.com/python/cpython/blob/6b7b9d00a9dd9cf66e3ec38b4c4b8a385b16b453/Python/import.c

With tests :
https://github.com/python/cpython/tree/6b7b9d00a9dd9cf66e3ec38b4c4b8a385b16b453/Lib/test/test_import
https://github.com/python/cpython/tree/6b7b9d00a9dd9cf66e3ec38b4c4b8a385b16b453/Lib/test/test_importlib