Report a clean error for corrupt cache instead of INTERNAL ERROR

[inchang-ing]

Summary

When the mypy cache is corrupt (for example, partially overwritten in a shared/temp directory), mypy used to crash with an INTERNAL ERROR. This PR replaces that with an actionable error telling the user to remove the cache directory (or pass a different --cache-dir) and re-run.

Fixes #22220

How it works

The fix adds guards on every cache-read path in mypy/build.py:

• Deserializing the cache data file (State.load_tree) — the most common corruption case (a garbage/trucated .data.ff).
• Resolving cache metadata (CacheMeta.read / CacheMeta.deserialize and CacheMetaEx).
• Fixing cross-references after deserialization (State.fix_cross_refs), where a missing type reference used to raise during fix-up.
• The deferred "TypeInfo not fixed" AssertionError (NOT_READY placeholder) that otherwise surfaced deep inside type checking, caught in State.wrap_context.

All paths now report a blocker error such as:

<cache>.data.ff: error: Error reading cache file; you likely have a corrupt cache.
Try removing the .mypy_cache directory and run mypy again.

Reproduction / verification

Before the fix, a corrupt cache produces:

error: INTERNAL ERROR -- Please try using mypy master on GitHub: ...

After the fix:

dep.data.ff: error: Error reading cache file; you likely have a corrupt cache.
Try removing the .mypy_cache directory and run mypy again.
Found 1 error in 1 file (errors prevented further checking)

A new regression test (mypy/test/testcache.py) populates the cache, corrupts the cached data file (preserving mtime so the cache is still trusted), and asserts the next build fails with a clean "corrupt cache" error and without an INTERNAL ERROR. The test fails on the unguarded code (raw AssertionError / INTERNAL ERROR) and passes with the fix.

Test plan

• python -m pytest mypy/test/testcache.py -n0 — new test passes.
• python -m pytest mypy/test/testmerge.py -n0 — existing cache/merge tests pass (no regression).

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

scipy-stubs (https://github.com/scipy/scipy-stubs)
+ tests/misc/test_submodules_empty.pyi:3: note: See https://mypy.readthedocs.io/en/stable/running_mypy.html#missing-imports