perf(parquet): decode DELTA_BYTE_ARRAY into contiguous storage

[fallintoplace]

Summary

• Decode DELTA_BYTE_ARRAY suffixes in one batch.
• Store reconstructed values in one contiguous byte arena per Decode call.
• Keep prefix validation before allocation and reconstruction.
• Keep separate arenas for separate decode calls so earlier results stay valid.
• Add coverage for partial reads, spaced reads, empty values, page resets, and invalid state.
• Add benchmarks for high and low prefix sharing with 1K and 64K values.

Benchmark

Apple M1 Pro. Compared with main at 6b039a76. The benchmark uses 65,536 values, 1 second per sample, 5 samples, and GOMAXPROCS=1.

case	main	branch	change	allocs/op
prefix-heavy, batch 128	3.678 ms	2.714 ms	26.2% faster	65,553 -> 529
prefix-heavy, full page	3.625 ms	2.727 ms	24.8% faster	65,551 -> 17
low-prefix, batch 128	3.172 ms	2.424 ms	23.6% faster	65,552 -> 529
low-prefix, full page	3.169 ms	2.382 ms	24.8% faster	65,551 -> 17

Command:

GOMAXPROCS=1 go test ./parquet/internal/encoding -run '^$' -bench '^BenchmarkDeltaByteArrayDecoderDecode/(prefix-heavy|low-prefix)/65536/batch-(128|65536)$' -benchmem -benchtime=1s -count=5 -cpu=1

Checks

• go test ./parquet/internal/encoding -count=1
• go test -race ./parquet/internal/encoding -count=1
• go vet ./parquet/internal/encoding
• go test ./parquet/... -run '^$'
• git diff --check