CUDA: add a fused top-K MoE kernel (#16130)

* CUDA: add a fused top-K MoE kernel

This kernel does the following:
1. softmax over the logits per token [n_experts, n_tokens]
2. argmax reduce over the top-k (n_experts_used) logits
3. write weights + ids to global memory

It is intended as fusion of softmax->top-k->get_rows pipeline for MoE models

* Refactor into ggml_cuda_should_use_topk_moe

* Review: Use better coalescing pattern, use WARP_SIZE, store logits into registers before

* Review: format + micro-optimizations

* Fix bug: fix tie breakers

* Add optional norm + clean-up code

* Use smem for final write

* Add bounds check

* Use better memory pattern for writeback

src/llama-graph.cpp

@@ -932,6 +932,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens]
     cb(weights, "ffn_moe_weights", il);
 
+
     if (gating_op == LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT) {
         weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);
         weights = ggml_soft_max(ctx0, weights); // [n_expert_used, n_tokens]
@@ -955,6 +956,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         cb(weights, "ffn_moe_weights_scaled", il);
     }
 
+    //call early so that topk-moe can be used
+    ggml_build_forward_expand(gf, weights);
+
     cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
 
     if (weight_before_ffn) {