Reduce per-worker memory of long-lived caches

[ondrejmirtes]

Two independent reductions of per-worker retained memory, found by profiling self-analysis (make phpstan reports the sum of per-worker peak memory):

1. Restore value sharing in name scope maps hydrated from the file cache

FileTypeMapper persists name scope maps with var_export(), which cannot represent shared references. A freshly created map shares the uses/constUses arrays and parent IntermediaryNameScope objects across all scopes of a file (COW / object references), but a map hydrated from the cache carries a private copy of everything in every scope — e.g. MutatingScope.php hydrates as ~200 scopes, each with its own copy of the same ~150-entry uses array and its own duplicate of the whole parent chain (~1.7 MB for that one file). Interning the hydrated scopes through a per-load pool restores the sharing: the retained size of FileTypeMapper::$memoryCache in an average worker drops from ~70 MB to ~7 MB.

2. Cap CachedParser by total cached source size

The AST of a parsed file takes roughly 50–60× more memory than its source text, and the parser cache limit was entry-count-based only — a few very large files (in self-analysis e.g. NodeScopeResolver.php, TypeCombinatorTest.php) each pinned tens of MB of AST in a worker for the whole run. The cache now also evicts by total source bytes of the cached entries (CACHED_SOURCE_BYTES_LIMIT, 512 KB of source ≈ ~30 MB of AST), on top of the existing count limit and LRU order. Eviction happens before insertion, so the entry being inserted is never evicted, and a single oversized file is still cached while it is being analysed.

Results (self-analysis, warm caches, 10-core machine / 8 workers):

	Used memory
2.2.x	2.72 GB
with this PR	2.30 GB

Per-worker peaks drop from ~340 MB to ~280 MB; elapsed time is unchanged. Full test suite passes.

🤖 Generated with Claude Code

https://claude.ai/code/session_017DjgMHgfvwhD4ogdTaRtLp