CCCL 3.4 benchmark on B200

[bernhardmgruber]

Basically same as #6728 but from the CCCL 3.4 branch.

CUB:

Thrust:

CUB in CCCL 3.2 vs 3.4:

Benchmark SQLite database added to the CCCL team drive:
https://drive.google.com/drive/folders/1eaRnrsz7gEmX386rgI0L_boRJIejNGDk

[bernhardmgruber]

Notes:

• Total benchmark runtime required 3 SLURM jobs, so was between 8h and 12h. That's annoying, we used to be able to finish with 2-3h for CCCL 3.0. Biggest offenders are again segmented algorithms.
• Many benchmarks were skipped by run.py and I don't know why (run.py did not run them):
  • thrust.is_partitioned.basic.base
  • thrust.is_sorted.basic.base
  • thrust.is_sorted.basic.with_predicate
  • thrust.is_sorted_until.basic.base
  • thrust.is_sorted_until.basic.with_predicate
  • thrust.mismatch.basic.base_range_iter
  • thrust.none_of.basic.base
  • cub.reduce.by_key.base
  • cub.segmented_scan.custom.fixed_size_segments
  • cub.segmented_scan.custom.varying_size_segments
  • cub.segmented_scan.sum.fixed_size_segments
  • cub.segmented_scan.sum.varying_size_segments
• It seems segmented reduction regressed between 3.2 and 3.4
• Two thrust.replace benchmarks are off the chart, but the results in the database look within bounds, so no idea what's going on there.