perf(arrow/array): reuse TableReader scratch slices

[fallintoplace]

What

TableReader.Next allocated two column slices for every record batch.

This keeps the batch slice on TableReader and reads the current chunks again after finding the batch size. NewRecordBatch copies the input slice, so the scratch slice can be safely reused on the next call.

Benchmark

The benchmark reuses one reader across passes. It uses Int32 columns with 256 rows per chunk.

256 columns x 256 chunks:

• 10.2 ms -> 3.0 ms
• 3.76 MB -> 1.27 MB
• 1024 -> 512 allocations

The benchmark covers 4, 32, and 256 columns with 32 and 256 chunks.

Tests

• go test ./arrow/array -count=1
• go test -race ./arrow/array -run TestTableReader -count=1