perf(parquet): SIMD BYTE_STREAM_SPLIT encoding

[fallintoplace]

What

• Add runtime-dispatched SIMD encoding for 4-byte and 8-byte BYTE_STREAM_SPLIT values.
• Use NEON structure loads and unzip operations on arm64.
• Use AVX2 unpacking on amd64.
• Reuse the fast path for numeric encoders and fixed-length byte arrays of width 4 or 8.
• Keep the scalar fallback for noasm builds and CPUs without the required feature.

Benchmark

Direct split benchmark with 65,536 values on an Apple M1 Pro, Go 1.26.3, 1 CPU, 150 ms, 7 runs. dispatch uses the NEON path on this machine.

Value width	Scalar	SIMD	Speedup
4 bytes	113,877 ns/op	6,745 ns/op	16.9x
8 bytes	218,824 ns/op	28,861 ns/op	7.6x

Tests

• go test ./parquet/internal/encoding
• go test -tags noasm ./parquet/internal/encoding
• GOOS=darwin GOARCH=amd64 go test -c -o /dev/null ./parquet/internal/encoding
• PARQUET_TEST_DATA="$(pwd)/parquet-testing/data" go test ./...
• go vet ./parquet/internal/encoding