[1/2] Add Kernel support for Cutlass based Fused FP4 MoE

[pavanimajety]

This kernel adds support for NVFP4 MoE kernels.

Currently measured perf:

[--------------------------------------------------------------------------------------------- FP4 MOE vs FP8 Triton ---------------------------------------------------------------------------------------------]
                                                                                                                       |  triton_moe  |  triton_moe_cuda_graphs  |  cutlass_moe_fp4  |  cutlass_moe_fp4_cuda_graphs
1 threads: --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
      nvidia/DeepSeek-R1-FP4, num_experts=256, topk=8, per_act_token=False per_out_ch=False, MKN=((4, 2048, 7168))     |     10.7     |            9.9           |        12.3       |              10.5
      nvidia/DeepSeek-R1-FP4, num_experts=256, topk=8, per_act_token=False per_out_ch=False, MKN=((8, 2048, 7168))     |     17.5     |           16.7           |        16.4       |              14.5
      nvidia/DeepSeek-R1-FP4, num_experts=256, topk=8, per_act_token=False per_out_ch=False, MKN=((16, 2048, 7168))    |     29.5     |           28.8           |        23.4       |              21.5
      nvidia/DeepSeek-R1-FP4, num_experts=256, topk=8, per_act_token=False per_out_ch=False, MKN=((32, 2048, 7168))    |     45.8     |           45.1           |        32.8       |              31.2
      nvidia/DeepSeek-R1-FP4, num_experts=256, topk=8, per_act_token=False per_out_ch=False, MKN=((64, 2048, 7168))    |     60.9     |           60.0           |        41.5       |              40.3
      nvidia/DeepSeek-R1-FP4, num_experts=256, topk=8, per_act_token=False per_out_ch=False, MKN=((128, 2048, 7168))   |     69.5     |           68.8           |        47.8       |              45.3
      nvidia/DeepSeek-R1-FP4, num_experts=256, topk=8, per_act_token=False per_out_ch=False, MKN=((256, 2048, 7168))   |     72.3     |           71.4           |        50.9       |              49.9
      nvidia/DeepSeek-R1-FP4, num_experts=256, topk=8, per_act_token=False per_out_ch=False, MKN=((512, 2048, 7168))   |     80.9     |           80.1           |        59.6       |              57.8
      nvidia/DeepSeek-R1-FP4, num_experts=256, topk=8, per_act_token=False per_out_ch=False, MKN=((1024, 2048, 7168))  |     93.7     |           92.8           |        75.9       |              74.7
      nvidia/DeepSeek-R1-FP4, num_experts=256, topk=8, per_act_token=False per_out_ch=False, MKN=((2048, 2048, 7168))  |    142.3     |          142.0           |       109.2       |             107.6

Times are in milliseconds (ms).

Follow up PR would include model support for nvidia/DeepSeekR1-FP4

Motivation

Modifications

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.