perf(parquet): decode nullable plain values into valid runs

[fallintoplace]

What does this change?

• PlainDecoder.DecodeSpaced used to decode all non-null values into a dense prefix first, then move them into their bitmap positions.
• For simple validity patterns, decode each valid run directly into its final output range.
• Keep the existing dense plus spaced expansion path for highly fragmented validity.
• Use the reverse bitmap reader so trailing-null pages stop after finding the dense prefix.
• Add edge-case tests and a focused benchmark.

Benchmark

Apple M1 Pro, Go 1.26.3, 65,536 int32 values, 8 runs, 300ms per sample:

Validity pattern	Before	After
leading null	11.6 us/op	7.1 us/op
clustered nulls	8.6 us/op	7.2 us/op
trailing null	5.6 us/op	5.6 us/op

Random 10% nulls and alternating validity keep the existing fallback path. All cases stayed at 0 B/op and 0 allocs/op.

Tests

• PARQUET_TEST_DATA=parquet-testing/data ARROW_TEST_DATA=arrow-testing/data go test -p 2 ./... -count=1
• go test -race -p 2 ./parquet/internal/encoding ./parquet/file
• go vet -p 2 ./parquet/internal/encoding ./parquet/file