perf(stats): persist per-file sizes in the on-disk index instead of reading lazily

[amondnet]

Context

Raised independently by the local code review on #92 and by Greptile on #92. Follows #90.

#92 reads local-path file sizes lazily at search time. That fixes the O(N) re-read on every cached load, but it leaves two gaps in file_chars telemetry:

• The size describes the file as it is now, not as it was when the returned chunk was indexed. During the MCP freshness cooldown, or with an explicitly loaded index whose source changed, saved_chars = file_chars - snippet_chars drifts. A file grown past MAX_FILE_BYTES since indexing contributes 0.
• A git index reloaded from cache (source_id is a URL) and a moved local source both report file_chars: 0 forever, because there is nothing to read lazily.

Proposal

Persist the per-file UTF-16 size map in the on-disk index at build time, alongside chunks.json / manifest.json, behind an INDEX_SCHEMA_VERSION bump so stale caches rebuild.

• save() writes the map captured at index time (the indexer already reads every file, so this is free).
• load_from_disk restores it; no search-time I/O at all.
• Collapses the FileSizes empty / captured / lazy split and the Mutex memo into a single captured map.
• Sizes then always match the indexed content, and git-from-cache and moved-source both work.

Notes

• Partial reindex (parity(#225): partial (incremental) reindexing — reuse unchanged files' chunks, vectors, and BM25 postings #84, ADR-0005) must update the persisted entries for changed files.
• Document the on-disk format change in .please/docs/references/semble.md and the relevant ADR.