perf(table): stream projected equality deletes

[fallintoplace]

Summary

• Read equality delete files with only the fields used by the delete key.
• Build encoded delete keys while record batches are being read.
• Keep the existing field ID and name mapping behavior.
• Preserve null handling, type encoding, cancellation, and the current worker limit.
• Add an equivalence test against the old materialized path.

Why

ReadTable loaded the full equality delete file before the keys were built. This also loaded columns that were not part of the equality key. Streaming the projected columns lowers peak memory use and avoids that extra work.

Checks

• go test ./table -count=1
• go test -race ./table -run "EqualityDelete|ReadEqualityDeleteFile" -count=1
• go vet ./table

Benchmark

For a 100k-row equality delete file on an Apple M1 Pro:

• Streamed projection: 30.6 MB/op, 11.6 ms/op
• Materialized table: 161.2 MB/op, 25.0 ms/op