Improve performance of FixedLengthBinary decoding

[etseidl]

Which issue does this PR close?

Tangentially related to #6219.

Rationale for this change

Following up on #6159 (comment), this adds set_from_bytes to the private ParquetValueType trait. This avoids the need for downcasting for byte array types.

What changes are included in this PR?

In addition to the above, it also changes the formatting in the encoding benchmark to distinguish between FixedLenByteArray(2) and FixedLenByteArray(16). Without this critcmp overwrites the FLBA(2) results.

Are there any user-facing changes?

No, changes are to private interfaces.

[etseidl]

While performance for decoding is still abominable, oddly this change improves it somewhat.

On my laptop (old macbook pro w/ 2.2GHz core i7)

% critcmp master set_from
group                                                                                master                                 set_from
-----                                                                                ------                                 --------
decoding: dtype=FixedLenByteArray(16), encoding=BYTE_STREAM_SPLIT                    1.09   564.8±15.15µs        ? ?/sec    1.00   517.2±10.01µs        ? ?/sec
decoding: dtype=FixedLenByteArray(2), encoding=BYTE_STREAM_SPLIT                     1.05    384.5±6.32µs        ? ?/sec    1.00    365.4±8.88µs        ? ?/sec
decoding: dtype=f32, encoding=BYTE_STREAM_SPLIT                                      1.01     38.8±0.86µs        ? ?/sec    1.00     38.6±0.85µs        ? ?/sec
decoding: dtype=f64, encoding=BYTE_STREAM_SPLIT                                      1.00     82.1±1.48µs        ? ?/sec    1.01     82.6±2.22µs        ? ?/sec
encoding: dtype=FixedLenByteArray(16), encoding=BYTE_STREAM_SPLIT                    1.02   282.4±15.78µs        ? ?/sec    1.00   277.1±14.91µs        ? ?/sec
encoding: dtype=FixedLenByteArray(2), encoding=BYTE_STREAM_SPLIT                     1.05     50.6±4.60µs        ? ?/sec    1.00     48.2±1.91µs        ? ?/sec
encoding: dtype=f32, encoding=BYTE_STREAM_SPLIT                                      1.00     44.4±1.42µs        ? ?/sec    1.01     44.9±1.69µs        ? ?/sec
encoding: dtype=f64, encoding=BYTE_STREAM_SPLIT                                      1.00    108.3±2.07µs        ? ?/sec    1.00    108.6±2.72µs        ? ?/sec

On my workstation (3.6 GHz Core i7-12700K)

group                                                                                master                                 set_from
-----                                                                                ------                                 --------
decoding: dtype=FixedLenByteArray(16), encoding=BYTE_STREAM_SPLIT                    1.19    357.0±2.69µs        ? ?/sec    1.00    299.7±1.98µs        ? ?/sec
decoding: dtype=FixedLenByteArray(2), encoding=BYTE_STREAM_SPLIT                     1.23    300.2±1.84µs        ? ?/sec    1.00    244.4±1.98µs        ? ?/sec
decoding: dtype=f32, encoding=BYTE_STREAM_SPLIT                                      1.00     15.1±0.31µs        ? ?/sec    1.00     15.1±0.13µs        ? ?/sec
decoding: dtype=f64, encoding=BYTE_STREAM_SPLIT                                      1.02     33.5±0.64µs        ? ?/sec    1.00     32.9±0.57µs        ? ?/sec
encoding: dtype=FixedLenByteArray(16), encoding=BYTE_STREAM_SPLIT                    1.01    120.8±2.49µs        ? ?/sec    1.00    119.6±0.52µs        ? ?/sec
encoding: dtype=FixedLenByteArray(2), encoding=BYTE_STREAM_SPLIT                     1.00     24.1±0.24µs        ? ?/sec    1.01     24.2±0.17µs        ? ?/sec
encoding: dtype=f32, encoding=BYTE_STREAM_SPLIT                                      1.01     16.5±0.20µs        ? ?/sec    1.00     16.4±0.12µs        ? ?/sec
encoding: dtype=f64, encoding=BYTE_STREAM_SPLIT                                      1.00     36.1±0.26µs        ? ?/sec    1.01     36.3±0.39µs        ? ?/sec

[alamb]

Thank you @etseidl -- even if this didn't improve performance in my opinion it would still be a valuable change simply for readability.