perf: bound indexing memory and flush partial progress on interrupt

[iggy]

Problem

gocate -updatedb was slow and memory-hungry on large trees. Profiling showed the cost was not hashing — it was the store write path:

• Every Upsert ran SELECT ... FirstRow() with no usable index → a full table scan per file (O(n²) total). modernc.org/ql ignores the index for FirstRow (verified: 136 ms/row at 200k rows, 1.0× with index).
• hashFile called os.ReadFile, so peak memory = file size.
• Each row committed in its own BEGIN TRANSACTION; … COMMIT;.

Result: on a 32-core box it sat at ~120% CPU because the hash workers finished fast but funneled into one consumer doing slow sequential scans.

Changes

• internal/index/hash.go — stream instead of slurp: imohash via SumSectionReader (fixed samples) and xxh3 via io.Copy into xxh3.New(). Peak memory is now bounded by the read buffer. The SectionReader is rewound between passes (it's left at the tail after sampling). Hashes are byte-identical to before, so existing DB rows stay valid.
• internal/index/index.go — RunCtx loads the host's rows once into a read-only snapshot shared by the walk (quick-skip) and consumer (insert-vs-update) with no locking, eliminating the per-row SELECTs. The consumer batches writes into single transactions (default 1000 rows / 4 GiB). The walk bounds concurrency by both file count (Workers) and bytes in flight (MaxBytes, default 256 MiB/worker) so a few huge files can't starve small-file parallelism.
• cmd/gocate/main.go — cancels the run's context on SIGINT/SIGTERM so the consumer flushes its in-flight batch (committing partial progress) before close; a second SIGINT exits 130.
• internal/store/store.go — LoadExisting (hostname-scoped) + WriteBatch (one transaction for many rows, rollback on error). Upsert reuses the batch path.

Tests

• TestHashFileMatchesReadFileBasis — proves streaming hashes equal the old os.ReadFile approach across small/medium/large files.
• TestLoadExistingScopesByHost — proves a multi-host DB can't leak cross-host rows into the path-keyed snapshot.
• All existing tests pass; go vet clean.

Verification (5.2 GB, 4530 files)

	wall	CPU	peak RSS
before	9.0 s	136%	~18 MB
after	3.7 s	~110%	~18 MB

And a SIGINT mid-run commits partial progress (verified: 20k rows after interrupt, DB reopens cleanly).