perf(parquet): decode RLE booleans into bitmaps

[fallintoplace]

What does this PR do?

• Writes repeated RLE boolean runs directly into the output bitmap
• Copies literal bit-packed runs directly without a []uint64 staging buffer
• Keeps partial reads and truncated-page errors aligned with the existing decoder

Why?

RleBooleanDecoder.DecodeToBitmap currently decodes through a [1024]uint64 buffer and writes each value one at a time. This change uses bulk bitmap fills for repeated runs and packed bitmap copies for literal runs.

Benchmarks

Apple M1 Pro, GOMAXPROCS=1, -benchmem, 1M values:

pattern	before	after	change
all true	2.65 ms	2.45 us	-99.9%
alternating	3.45 ms	0.20 ms	-94.2%
short runs	3.50 ms	0.20 ms	-94.3%

The benchmark also covers 64K values, unaligned output offsets, and nullable spaced decoding.

Tests

• go test ./parquet/internal/encoding ./parquet/internal/utils ./parquet/pqarrow
• go test -race ./parquet/internal/encoding ./parquet/internal/utils ./parquet/pqarrow
• go test -tags noasm ./parquet/internal/encoding ./parquet/internal/utils

[zeroshade]

Just needs the merge conflicts to get resolved