[BUG] Improve performance of mixed joins on H100

[GregoryKimball]

When comparing x86-H100 versus x86-V100 microbenchmark performance in libcudf, we found that mixed join benchmarks showed slower runtimes on H100. The rest of libcudf microbenchmarks tend to be around 2-3x faster. Perhaps we need to adjust DEFAULT_JOIN_BLOCK_SIZE (code pointer), or some other performance hinting on the mixed join kernels.

Also see #10534 which added launch bounds to mixed joins kernels.

Figure showing H100 vs V100 speedup results:

Figure zooming in on JOIN benchmarks:

[Based on libcudf 23.08 commit aed7174eae6c6]

[GregoryKimball]

I reran the JOIN_NVBENCH benchmarks on V100, A100, H100 with 23.10 commit 9e099cef25b11. I could not reproduce the issue. I suspect the original observations were not using the same commit.