chore/gen_pypi_cache overhaul

[blast-hardcheese]

Why

Statically analyzing which modules are provided by a given package can give us some serious speed improvements.

Pseudoscientific performance graph of execution time before/after this PR:

Where	Time	Graph	Enjoyment
Current HEAD	1h30m	📈	📉
With this PR	16m	📉	📈

Additionally, looking at some user reports, relying on pkgutil seems to miss packages that are valid. As an example:

$ mkdir foo/bar/baz
$ touch foo/bar/baz/blix.py
$ python -c 'import foo.bar.baz.blix'

exits cleanly, though pkgutil would fail to see inside foo due to foo not being a module itself.

A real-world example of no longer overreporting modules:

-{"name":"3scale-api",...,"modules":["certifi","charset_normalizer","idna","requests","threescale_api","urllib3"]}
+{"name":"3scale-api",...,"modules":["threescale_api"]}

-{"name":"ailever",...,"modules":["FinanceDataReader","PIL","_brotli","_plotly_future_","_plotly_utils","ailever","brotli","bs4","certifi","charset_normalizer","click","cycler","dash","dash_bootstrap_components","dash_core_components","dash_html_components","dash_table","dateutil","flask","flask_compress","fontTools","idna","importlib_metadata","itsdangerous","jinja2","jupyterlab_plotly","kiwisolver","lxml","markupsafe","matplotlib","numpy","packaging","pandas","patsy","plotly","pylab","pyparsing","pytz","requests","requests_file","requests_futures","scipy","seaborn","six","soupsieve","statsmodels","tenacity","tqdm","urllib3","werkzeug","yahooquery","zipp"]}
+{"name":"ailever",...,"modules":["ailever"]}

What changed

• Add test-one subcommand for easy manual testing
• Rely on the python runtime to determine which file extensions are valid for considering whether they are modules or not
• Use pip install --no-deps to limit search blast radius. Important for "extension"-style modules which depend on the underlying package, which would obscure which package supplies the extension (flaskext.mysql, as an example, may also falsely report flask as a provided module)

Test plan

$ go run ./gen_pypi_map test -force

observe that it is faster

$ go run ./gen_pypi_map test-one -force flask-mysql
...

observe output

[airportyh]

nice! no deps makes this fast.

[blast-hardcheese]

I added unconditional retry here to try and catch transient failures, it added a few minutes more to execution time but not much more

[airportyh]

Exciting! I am testing this...

[airportyh]

lgtm