sm121 (GB10): follow-up measurements offered after #2039

[yashb98]

Follow-up to #2039, cc @matthewdouglas.

In the review of #2030 you mentioned a layer of dispatch decisions you couldn't test on real hardware. I have a GB10 (sm121) permanently available. Two things I can measure and report:

1. The six M in {4, 6, 8} shapes at N=2048, K=2048 that still pick SIMT where MMA measures faster.
2. The sub-wave M=4 region that highbw_gddr would affect if it were extended to sm121.

Would either be useful, and is there a release you'd want sm121 numbers for before it ships?