perf(table): avoid quadratic snapshot-log pruning

[fallintoplace]

What changed

• Build one snapshot-ID set in RemoveSnapshots.
• Reuse it for snapshot, snapshot-log, statistics, and partition-statistics pruning.
• Add a benchmark covering larger snapshot logs.

Benchmark

Same benchmark, 5 runs, 500 ms each, on an Apple M1 Pro:

• 1,000 entries: ~167 µs -> ~15.7 µs, about 10x faster.
• 10,000 entries: ~13.4 ms -> ~171 µs, about 78x faster.

The speedup comes from replacing repeated linear membership scans with map lookups.