[Kernel] add custom moe ops for prefill (#4194)

### What this PR does / why we need it?
1.Add the implementation of normal Aclnn operators: MoeCombineNormal,
MoeDispatchNormal, NotifyDispatch，and DispatchLayout.

- MoeCombineNormal: Implements the combine logic within MoE operations.
- MoeDispatchNormal: Implements the dispatch logic within MoE
operations.
- NotifyDispatch: Exchanges topk_idx information among different ranks
to calculate the device memory required for the dispatch stage.
- DispatchLayout: Used to calculate information related to the device
memory layout for the dispatch stage.

2.Provide PyTorch interfaces for normal operators—get_dispatch_layout,
dispatch_prefill, and combine_prefill—to be used for MoE communication
during the prefill stage in vLLM.

- get_dispatch_layout: Calculates information related to the device
memory layout for the dispatch operator, and is called before
dispatch_prefill.
- dispatch_prefill: Initiates the dispatch operation.
- combine_prefill: Initiates the combine operation.

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
The functionality has already been validated using the local Qwen model.
Test cases will be added after support for multi-NPU use cases in the CI
pipeline is finalized.


- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: shiro-zzzz <zhangdianhao@huawei.com>

csrc/notify_dispatch/op_kernel/notify_dispatch.cpp

@@ -0,0 +1,57 @@
+#include "kernel_operator.h"
+#include "notify_dispatch.h"
+#include "notify_dispatch_tiling.h"
+
+#define TILING_KEY_FLOAT16 20
+#define TILING_KEY_BFLOAT16 21
+#define TILING_KEY_FLOAT 22
+#define TILING_KEY_INT 23
+
+#define KERNEL_USE_WORKSPACE (1 * 1024 * 1024)
+
+extern "C" __global__ __aicore__ void notify_dispatch(
+    GM_ADDR sendData, GM_ADDR tokenPerExpertData, GM_ADDR sendDataOffset, GM_ADDR recvData, GM_ADDR workspace, GM_ADDR tiling)
+{
+    REGISTER_TILING_DEFAULT(NotifyDispatchTilingData);
+    GET_TILING_DATA_WITH_STRUCT(NotifyDispatchTilingData, tilingData, tiling);
+
+    // hcomm will set magic later in init
+    uint32_t magic = 1;
+    GM_ADDR commArgs = nullptr;
+
+    int localRank = tilingData.notifyDispatchInfo.localRankId;
+    int localRankSize = tilingData.notifyDispatchInfo.localRankSize;
+    int rank = tilingData.notifyDispatchInfo.rankId;
+    int rankSize = tilingData.notifyDispatchInfo.rankSize;
+    int64_t len = tilingData.notifyDispatchInfo.sendCount;
+    int64_t numTokens = tilingData.notifyDispatchInfo.numTokens;
+
+    GM_ADDR sendDataInput = sendData;
+    GM_ADDR tokenPerExpertDataInput = tokenPerExpertData;
+    GM_ADDR sendDataOffsetOutput = sendDataOffset;
+    GM_ADDR recvDataOutput = recvData;
+
+    // fill in unused args
+    uint32_t extraFlag = 0;
+    GM_ADDR scale = nullptr;
+    int root = 0;
+    int op = 0;
+    int cycleCount = 0;
+    int64_t scaleCount = 0;
+    GM_ADDR offset = nullptr;
+    int blockNum = GetBlockNum();
+
+    if (TILING_KEY_IS(TILING_KEY_FLOAT16)) {
+        NotifyDispatch<float16_t> opKernel(rank, rankSize, extraFlag);
+        opKernel.Init(KERNELS_ARGS_CALL_ALL2ALL());
+        opKernel.Process();
+    } else if (TILING_KEY_IS(TILING_KEY_FLOAT)) {
+        NotifyDispatch<float> opKernel(rank, rankSize, extraFlag);
+        opKernel.Init(KERNELS_ARGS_CALL_ALL2ALL());
+        opKernel.Process();
+    } else if (TILING_KEY_IS(TILING_KEY_INT)) {
+        NotifyDispatch<int> opKernel(rank, rankSize, extraFlag);
+        opKernel.Init(KERNELS_ARGS_CALL_ALL2ALL());
+        opKernel.Process();
+    }
+}