Add moe topk softmax templated from vllm (#4302)

sgl-kernel/include/sgl_kernel_ops.h

@@ -173,6 +173,12 @@ void moe_align_block_size(
     torch::Tensor token_cnts_buffer,
     torch::Tensor cumsum_buffer);
 
+void topk_softmax(
+    torch::Tensor& topk_weights,
+    torch::Tensor& topk_indices,
+    torch::Tensor& token_expert_indices,
+    torch::Tensor& gating_output);
+
 /*
  * From csrc/speculative
  */