Extend support for BYTE_STREAM_SPLIT to FIXED_LEN_BYTE_ARRAY, INT32, and INT64 primitive types

[anjakefala]

Please correct me if I'm wrong! It seems arrow-rs has added BYTE_STREAM_SPLIT support for float types, but not for other numerical data types like INT32.

Since then, the Parquet spec has been expanded to extend BYTE_STREAM_SPLIT encoding to other numerical primitive types: apache/parquet-format#229. The C++ PoC is here: apache/arrow#40094.

It would be good for arrow-rs to support BYTE_STREAM_SPLIT encoding additionally for FIXED_LEN_BYTE_ARRAY, INT32, and INT64.

[etseidl]

@anjakefala are you working on this? If not I have some free cycles to devote to this.

[anjakefala]

@etseidl I am not! I would love it if you picked it up. =)

[anjakefala]

Thank you so much @etseidl and @alamb! =)

[alamb]

Thank you so much @etseidl and @alamb! =)

Also @mapleFU 🙏

[etseidl]

@anjakefala please let me know if you have performance issues, particularly with FIXED_LEN_BYTE_ARRAY data. That was particularly problematic on the read side. 😦

[alamb]

label_issue.py automatically added labels {'parquet'} from #6159