Support --package-root with native parser

[ilevkivskyi]

When --package-root is used, fscache creates some fake __init__.py files on-the-fly (that don't actually exist). These obviously can't be handed by the native parser, but fscache.exists() returns True for them, causing a crash. One possible fix is to simply use os.path.isfile() everywhere. But I think fscache can give us a little perf boost, so I instead add a real_only flag to fscache.exists().

cc @JukkaL

[JukkaL]

The way the fake package cache interacts with the native parser is pretty sketchy and we may want to do it differently eventually, but this is fine as a quick fix.

[JukkaL]

os.path.dirname(path) may be faster, as it may construct one fewer substring (but haven't verified).

[ilevkivskyi]

OoC I checked and it is indeed ~20% faster.

[ilevkivskyi]

The way the fake package cache interacts with the native parser is pretty sketchy and we may want to do it differently eventually, but this is fine as a quick fix.

Yeah, --package-root is one of those dark corners I didn't risk touching. But also we will hopefully have string parsing support in native parser soon, see mypyc/ast_serialize#54. When we will have that, and the old parser will be deleted, we will be able to simplify a bunch of things.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅