In particular, CUDA 12 removes support for Kepler (3.x) entirely.
Unusually, however, it does not introduce any new architectures,
or even compute capabilities.
It turns out we don't generally need to proxy every compiler ever
through the top-level package. The number of times we directly poke at
one is negligible and direct imports are pretty clean.
Thanks to `ModuleInfo`, all modules are just named `foo.py` instead of
`unstable_foo.py`, which simplifies the import method a bit. This also
allows for accurate FeatureNew/FeatureDeprecated use, as we know when
the module was added and if/when it was stabilized.
David Seifert