Add fused top-K softmax kernel for MoE (#2769)

csrc/moe/moe_ops.cpp

@@ -0,0 +1,7 @@
+#include "moe_ops.h"
+
+#include <torch/extension.h>
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("topk_softmax", &topk_softmax, "Apply topk softmax to the gating outputs.");
+}