perf(parquet): reuse DELTA decoder scratch

[fallintoplace]

What

• Reuse DELTA length and prefix scratch buffers across decoder pages.
• Add tests for scratch reuse and page-to-page correctness.

Why

• The column reader caches decoders by encoding.
• Each page currently allocates fresh length and prefix arrays even when the decoder already has enough capacity.

Implementation

• Keep stable backing buffers for decoded lengths and prefix lengths.
• Expose sliced views to the existing Decode and Discard paths.
• Keep the existing page-size behavior when a larger page needs more capacity.
• The existing repeated SetData + Decode benchmark now shows 18 to 15 allocations/op and about 512 KiB less allocation for a full 65,536-value prefix-heavy page.

Tests:

• go test ./parquet/internal/encoding -count=1
• go test ./parquet/file -run '^(TestWithEOFReader|TestInvalidHeaders|TestInvalidFooter|TestIncompleteMetadata|TestDeltaLengthByteArrayPackingWithNulls|TestDeltaBinaryPackedMultipleBatches|TestPageStreaming.*|TestPrimitiveReader|TestFullSeekRow|TestSkipEmptyRepeatedRows)$' -count=1
• go test -race ./parquet/internal/encoding -run 'TestDelta(ByteArrayDecoderReusesPageScratch|LengthByteArrayDecoderReusesLengthScratch|ByteArrayDecoderKeepsResultsAcrossPages|ByteArrayDecoderRejectsInvalidPrefixes)$' -count=1
• go vet ./parquet/internal/encoding

[zeroshade]

The scratch ownership and reset behavior look sound. lengthScratch and prefixScratch are decoder-owned metadata storage, each SetData re-slices from index zero, and decoded byte-array results do not alias either scratch buffer. Page reuse, partial consumption, malformed counts, and growth paths preserve the existing behavior.