Improve parquet reader very-long string performance

[pmattione-nvidia]

The previous strings PR significantly reduced the parquet reader string performance for very-long strings, for lengths ~1024 and longer. This PR fixes the performance issue by instituting a max memcpy length of 8 bytes at once (this length yielded best perf). Also, up to all of the threads in the block can work on the same string, rather than limiting it to just all of the threads in a warp.

PERFORMANCE:
Short strings: Unchanged
Length 1024: 25% faster
Longer lengths (up to 64k): Up to 90% faster, same as before strings PR

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[vuule]

Could you post the impact of the change on the benchmarks? Not required to merge IMO, but it's nice to keep such result available long-term.

[vuule]

Looks good. I really like the simplification in calc_threads_per_string_log2.

[pmattione-nvidia]

Could you post the impact of the change on the benchmarks? Not required to merge IMO, but it's nice to keep such result available long-term.

Done

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[galipremsagar]

@pmattione-nvidia I cancelled the most recent workflow to free up resources to unblock all of cudf CI for this PR: #17771

I'll rerun once #17771 is merged.

[pmattione-nvidia]

/merge