[ENH] Speed up first completion

[peytondmurray]

This PR speeds up the first completion by

• Removing some unnecessary list allocations
• Pushing iteration over directory contents into compiled code. This avoids calling os.scandir, os.path.basename, and inspect.getmodulename on every directory entry.
• An import caching system now tracks the state of any paths targeted by python's import machinery. Upon first completion, the modules available on each path are read using the new compiled extension before being stored in <user cache dir>/pyflyby/. Any subsequent completions (including in other ipython sessions) check the mtime of the path, and look for a corresponding file in <user cache dir>/pyflyby/. If a valid cache file is found, module names used for completion are read from the file rather than from the filesystem. Otherwise, the cache is rebuilt.
• A function to force all import caches to be rebuilt is now available at pyflyby.rebuild_import_cache.

Other changes:

• The build backend is now meson-python>=0.18.0
• etc/ and libexec/ are now placed in <site-packages>/pyflyby instead of polluting other parts of the filesystem.
  I also removed some unused python 2 compatibility stuff adjacent to the code I was working on - if it's better to leave this for another PR I'll revert it.
• appdirs and prompt_toolkit are now dependencies of pyflyby.

Closes #571.

[krassowski]

During installation I encountered an error coming from mesonpy:

if licence_ and license.file:
                ^^^^^^^^^^^^
AttributeError: 'str' object has no attribute 'file'. Did you mean: 'title'?

It looks like the mesonpy version I get expects the old format (https://packaging.python.org/en/latest/guides/writing-pyproject-toml/#license).

I believe a lower bound on required mesonpy would be needed, or change in pyproject.toml to use the old format.

More details:

• meson-python docs still uses the old format
• the new format support was added in v0.18.0 in ENH: add support for PEP 639 mesonbuild/meson-python#681

[peytondmurray]

Note that due to the additional log messages this breaks some tests which check ipython output. I'll address these if the performance gains here are sufficient to warrant this change.

[Carreau]

@peytondmurray do you remember if this is a leftover from local testing ? I can't find any use of this.

[peytondmurray]

It's been a while since I've touched this, but it's possible that the request for import caching also included a way to rebuild the import cache manually. If it's not being used anywhere, I think it's probably fine to remove it.

[krassowski]

Yes, I think this is the relevant context:

Of course for cases where file contents were modified but no files were added or removed, this wouldn't regenerate the cache like we'd really like. But this may be an edge case that we can handle by adding a function to manually regenerate the cache anyway.

So, this function is probably needed for manually refreshing cache as an escape hatch and should be documented/exposed in public API (or removed if this is rare corner case)