perf(table): flatten manifest entries after parallel reads

[fallintoplace]

Summary

• Keep one classified result per manifest while reads run in parallel.
• Flatten after all reads finish into exact-sized data and delete slices.
• Preserve manifest-list order and keep manifestEntries.merge for existing callers.
• Add focused tests for mixed buckets, empty result slots, and exact capacities.
• Add a merge-versus-flatten benchmark.

Benchmark

Apple M1 Pro. 64 manifests, 10,000 entries per manifest, concurrency 16.

• Data manifests: 18.4 ms/op and 71.9 MB/op with merge, versus 6.4 ms/op and 20.7 MB/op with flatten.
• Delete manifests: 12.9 ms/op and 74.6 MB/op with merge, versus 6.3 ms/op and 32.4 MB/op with flatten.

Tests

• go test ./... -count=1
• go vet ./table
• Targeted go test -race ./table pass.