arrow

GH-39122: [C++][Parquet] Optimize FLBA record reader (#39124)

### Rationale for this change

The FLBA implementation of RecordReader is suboptimal:
* it doesn't preallocate the output array
* it reads the decoded validity bitmap one bit at a time and recreates it, one bit at a time

### What changes are included in this PR?

Optimize the FLBA implementation of RecordReader so as to avoid the aforementioned inefficiencies.
I did a quick-and-dirty benchmark on a Parquet file with two columns:
* column 1: uncompressed, PLAIN-encoded, FLBA<3> with no nulls
* column 2: uncompressed, PLAIN-encoded, FLBA<3> with 25% nulls

With git main, the file can be read at 465 MB/s. With this PR, the file can be read at 700 MB/s.

### Are these changes tested?

Yes.

### Are there any user-facing changes?

No.
* Closes: #39122

Lead-authored-by: Antoine Pitrou <antoine@python.org>
Co-authored-by: Antoine Pitrou <pitrou@free.fr>
Signed-off-by: Antoine Pitrou <antoine@python.org>