Add batched NVFP4 MoE GEMM for SM_100 (B200) with benchmark

Batched CUTLASS GEMM using rank-4 problem shape (M,N,K,L) for MoE
inference. Fixed max_M per expert with zero-padded activations,
CUDA-graph-friendly design. Achieves 2.0-2.4x speedup over BF16
cuBLAS on B200 across GLM-4.7 MoE shapes (8 experts, 8-128
tokens/expert). Peak 1940 TFLOPS NVFP4 vs 830 TFLOPS BF16.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: TimDettmers

CMakeLists.txt

@@ -301,6 +301,7 @@ if(BUILD_CUDA)
             set(_NVFP4_SM100_SOURCES
                 csrc/qutlass/gemm_nvfp4_sm100.cu
                 csrc/qutlass/gemm_nvfp4_grouped_sm100.cu
+                csrc/qutlass/gemm_nvfp4_moe_sm100.cu
             )
 
             add_library(nvfp4_sm100a OBJECT ${_NVFP4_SM100_SOURCES})