perf(arrow/compute): avoid goroutines for serial record filtering

[fallintoplace]

Summary

• Run FilterRecordBatch columns synchronously for serial execution.
• One-column batches use the direct path even when parallelism is configured.
• NumParallel <= 1 uses the direct path for all columns.
• Keep the errgroup path for multi-column parallel execution.
• Add coverage for one-column, zero/one worker, parallel, null selection, and length errors.

Benchmark

Local Apple M1 Pro run with NumParallel=1:

• 8 columns / 16 rows: 53.3 us -> 26.2 us, 263 -> 240 allocs/op
• 128 columns / 16 rows: 709.8 us -> 420.9 us, 3,983 -> 3,720 allocs/op
• 128 columns / 4,096 rows: 1.07 ms -> 0.77 ms, 4,046 -> 3,783 allocs/op

Tests

• go test ./...
• GOOS=linux GOARCH=386 go build ./arrow/compute/...