perf(table): cache snapshot manifests across scans

[fallintoplace]

What

• Cache decoded manifest lists per table and reuse them across fresh scans.
• Share concurrent first reads so only one manifest-list read runs for a snapshot.
• Keep data and delete manifest groups ready for planning paths that need one type.
• Reset the cache on Refresh so a new metadata snapshot never reuses stale descriptors.
• Apply the cache to local scans, incremental append scans, table inspection, and AllManifests.

Benchmark

Apple M1 Pro, 1 manifest, 10 repeated scans per benchmark iteration:

• Manifest-list opens: 10 → 1
• Time: 4.26 ms → 2.36 ms
• Memory: 4.74 MB → 2.78 MB
• Allocations: 37,048 → 23,519

With 10,000 manifests, manifest-list opens still drop from 10 to 1. Total planning time stays close because opening the 10,000 manifest files dominates that workload.

Checks

• go test ./table/... -count=1
• go test -race ./table/... -count=1
• go test ./... -run ^$ -count=1
• go vet ./...
• golangci-lint run --timeout=10m

🔥