refactor: 统一硬件相关头文件引用

将分散在各文件中的CUDA/HIP/MUSA硬件相关头文件引用统一到vendors目录下的对应头文件中，提高代码可维护性。移除重复的头文件引用，优化构建配置。

csrc/moe/permute_unpermute_kernels/dispatch.h

@@ -1,5 +1,5 @@
 #pragma once
-#include <cuda_fp8.h>
+#include "vendors/functions.h"
 #define MOE_SWITCH(TYPE, ...)                                     \
   at::ScalarType _st = ::detail::scalar_type(TYPE);               \
   switch (_st) {                                                  \

csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.h

@@ -2,12 +2,11 @@
 // reference from tensorrt_llm moe kernel implementation archive in
 // https://github.com/BBuf/tensorrt-llm-moe/tree/master
 
-#include <c10/core/ScalarType.h>
 #include <torch/all.h>
 #include "dispatch.h"
-#include <cub/cub.cuh>
-#include <cub/device/device_radix_sort.cuh>
-#include <cub/util_type.cuh>
+
+
+#include "../../vendors/functions.h"
 #include "cutlass/numeric_size.h"
 #include "cutlass/array.h"