Tune SM100 GEMM scheduling for dW shapes

[justinhh4]

Summary

• keep the fixed two-CTA SM100 GEMM design
• use swizzle-2 for measured large-M, small-K dW shapes
• select 4 A/B stages for K=256/2048 and 6 stages for K=512/1024
• retain the existing 6-stage, swizzle-1 configuration for all other shapes

B200 BF16 results

Representative M=128256, N=4096 results (median latency):

K	Before	After	Speedup	cuBLAS
256	0.353 ms	0.252 ms	1.40x	0.264 ms
512	0.511 ms	0.394 ms	1.30x	0.396 ms
1024	0.932 ms	0.699 ms	1.33x	0.685 ms
2048	1.779 ms	1.349 ms	1.32x	1.336 ms

Across 16 Llama, Qwen, and DeepSeek dW shapes, the tuned configuration was 1.07-1.40x faster than the previous configuration and stayed within about 4% of cuBLAS, often slightly faster.

Testing

• python -m pytest test/cutedsl/test_sm100_gemm.py -q (11 passed)
• ruff check and ruff format --check on changed files