perf(table): precompute data writer invariants

[fallintoplace]

What changed

• Precompute the stats plan, current partition spec, Parquet row-group size, write properties, schema options, and file extension when the writer is created.
• Reuse those values for every write task.
• Add a focused writer benchmark.

Benchmark

Apple M1 Pro, GOMAXPROCS=1, median of 5 runs with -benchtime=1s:

• origin/main: 8.53 us/op, 97 allocs/op, 6.49 KB/op
• this branch: 6.38 us/op, 72 allocs/op, 5.01 KB/op
• about 1.34x faster and 26% fewer allocations

The benchmark uses a no-op file format so it measures the per-task writer setup and schema conversion without filesystem or Parquet I/O.

Tests

• go test ./...
• go test -race ./table/...
• go vet ./table/...

[zeroshade]

Since this is a draft, i'll hold off on further review until it's marked ready