perf(arrow/compute): take chunked binary values without concatenation

[fallintoplace]

What

Chunked string and binary take currently concatenates all value chunks before selecting rows.

This adds a chunk-aware path for:

• STRING
• BINARY
• LARGE_STRING
• LARGE_BINARY

The new path resolves global indices to source chunks and copies only selected values. It keeps one output chunk per index chunk and preserves null and bounds behavior.

Benchmark

64 chunks x 4096 rows, 32-byte values, random indices, Apple M1 Pro:

• 1% selected: 0.87 ms / 9.6 MB -> 0.08 ms / 0.11 MB
• 10% selected: 2.36 ms / 10.4 MB -> 1.43 ms / 0.96 MB

The benchmark covers 8 and 64 chunks, 1% / 10% / 50% / 100% selection, and all four types.

Tests

• go test ./arrow/compute/... -count=1