Fix CUDA RMSNorm small-row dispatch

[dhiltgen]

Proposed changes

CUDA RMSNorm used a launch shape that packed multiple rows into one block even when each row required more than one reduction group. BlockBroadcastReduce is block-wide, so those rows mixed partial reductions and produced incorrect results for small dimensions such as fp32 D=256 and bf16/fp16 D=512.

Fix the affected dispatch case by launching one row per block when two reduction groups are needed. Add CUDA regression coverage for forward and VJP fast RMSNorm on the affected small multi-row shapes.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[zcbenz]

Nice catch, thanks!