Re-resolve ShortNameResolver short names on stmts change

[TomasVotruba]

What

ShortNameResolver::resolveFromFile() cached its short-name-to-FQN map by file path and never invalidated it. Across the file-processing convergence loop the stmts change between iterations, so later iterations read a stale map - a correctness bug - and the map for every processed file was retained for the whole run.

Change

resolveFromFile() now caches by file identity + getNewStmtsVersion() and re-resolves when the stmts version bumps - the same pattern resolveShortClassLikeNames() already uses in this class. The per-file-path array is dropped, so only the current file's map is held.

Proof the bug is real

ShortNameResolverTest::testResolveFromFileAgainAfterStmtsChange resolves a file, renames its class, bumps the stmts version, and resolves again. It mirrors the existing testResolveShortClassLikeNamesAgainAfterStmtsChange guard for the sibling method.

• on the old filePath cache it fails: Failed asserting that an array does not have the key 'VariousImports' (stale map returned after the change)
• with this fix it passes

Tradeoff - correctness + memory, costs some CPU

Measured with bench/short_name_resolver_bench.php (8 convergence passes, 400 node queries each, 25 repeats):

pattern	before (filePath cache)	after (version cache)
single-version	0.51s cpu	0.58s cpu
multi-version (stmts bumped each pass)	0.50s cpu	1.05s cpu

The old cache ignored version bumps, so convergence passes were free; the new one re-resolves per version (~2x cpu on the multi-version pattern). The cost only hits files that actually change across iterations, scaling with iteration count. In return the short-name map is correct on every iteration and only one file's map is held at a time.

Tests

148 import tests green, plus the new regression test. ECS + full PHPStan clean. Benchmark lives in bench/ (not scanned by ECS/PHPStan/Rector).