Summary

This fixes a few bugs introduced by #5104. I previously thought we could track conflicting locals the same way we track conflicting URLs in forks, but it turns out that ends up being very tricky. URL forks work because we prioritize directly URL requirements. We can't prioritize locals in the same way without conflicting with the URL prioritization (this may be possible but it's not trivial), so we run into issues where a correct resolution depends on the order in which dependencies are traversed.

Instead, we track local versions across all forks in Locals. When applying a local version, we apply all locals with markers that intersect with the current fork. This way we end up applying some local versions without creating a fork. For example, given:

// pyproject.toml
dependencies = [
    "torch==2.0.0+cu118 ; platform_machine == 'x86_64'",
]

// requirements.in
torch==2.0.0
.

We choose 2.0.0+cu118 in all cases. However, if a disjoint fork is created based on local versions, the resolver will choose the most compatible local when it narrows to a specific fork. Thus we correctly respect local versions when forking:

// pyproject.toml
dependencies = [
    "torch==2.0.0+cu118 ; platform_machine == 'x86_64'",
    "torch==2.0.0+cpu ; platform_machine != 'x86_64'"
]

// requirements.in
torch==2.0.0
.

We should also be able to use a similar strategy for #5150.

Test Plan

This fixes #5220 locally for me, as well as a few other bugs that were not reported yet.