perf(parquet): batch plain BYTE_ARRAY encoding

[fallintoplace]

Rationale for this change

PlainByteArrayEncoder.Put currently calls PutByteArray for every value. Each call checks sink capacity, writes the 4-byte length, then writes the payload. Large batches repeat that sink work for every value.

What changes are included in this PR?

• Calculate the encoded batch size in one pass.
• Reserve the sink capacity once.
• Write lengths and payloads directly into the reserved region.
• Add benchmarks for Put and all-valid PutSpaced with 1,048,576 values at three widths.

The single-value PutByteArray path is unchanged. PutSpaced benefits through its existing valid-run batching.

Medians from 6 runs on an Apple M1 Pro were:

method	value width	main	this PR	change
Put	4 bytes	10.25 ms	3.85 ms	-62.4%
Put	16 bytes	11.00 ms	4.38 ms	-60.2%
Put	64 bytes	13.77 ms	7.00 ms	-49.2%
PutSpaced	4 bytes	10.33 ms	3.95 ms	-61.8%
PutSpaced	16 bytes	11.13 ms	4.44 ms	-60.1%
PutSpaced	64 bytes	13.54 ms	7.45 ms	-45.0%

Allocation counts are unchanged.

Are these changes tested?

• go test ./parquet/... -count=1
• go test ./parquet/internal/encoding -run "^$" -bench "^BenchmarkEncodePlainByteArray$" -benchmem -count=6

The existing encoding suite covers plain BYTE_ARRAY round trips and spaced validity patterns.

Are there any user-facing changes?

No. The encoded Parquet bytes and public API are unchanged.