Use parallel parsing at all stages

[ilevkivskyi]

This is a follow-up to #21175 and #21119.

Right now we are only using parallel parsing during initial graph loading. This PR allows using it when processing SCCs that are stale due to dependencies (these are not parsed on initial loading). Implementation is a bit less trivial than I thought, because we need to completely extract the tree de-serialization (which is not parallelizeable). It is still relatively straightforward. The new logic is like this:

• When calling parse(eager=False) (default) we always return a serialized tree.
• When calling parse(eager=True) a caller can force immediate de-serialization (when suitable), this will trigger load_from_raw().
• To get more manual control (e.g. for imports_only=True) one can use parse(eager=False) followed by load_from_raw() directly.

cc @JukkaL

[github-actions]

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