perf(parquet/compress): reuse Brotli writers across encodes

[fallintoplace]

Summary

• Reuse Brotli writers for one-shot Encode and EncodeLevel calls.
• Keep a separate sync.Pool for each supported compression level.
• Reset pooled writers with a nil destination before putting them back.
• Keep the streaming writer APIs unchanged.
• Add serial and parallel benchmarks plus sequential and concurrent correctness tests.

Benchmark

Compared with main at 6b039a76:

Case	Main	This PR	Main allocations	This PR allocations
compressible / 256 KiB	~0.78 ms/op	~0.51 ms/op	~22 allocs/op, ~11.9 MB/op	1 alloc/op, ~10 KB/op
semi-random / 256 KiB	~2.04 ms/op	~1.77 ms/op	~21 allocs/op, ~12.0 MB/op	1 alloc/op, ~37 KB/op

The parallel benchmark also stays at one allocation per operation in the pooled path.

Checks

• go test ./parquet/compress -count=1
• go test -race ./parquet/compress
• go vet ./parquet/compress
• Parquet package tests with PARQUET_TEST_DATA set