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/quantization/w8a8/cutlass/c3x/cutlass_gemm_caller.cuh

@@ -2,9 +2,9 @@
 
 // clang-format will break include orders
 // clang-format off
-#include <torch/all.h>
+#include "../../../../vendors/functions.h"
+
 
-#include <ATen/cuda/CUDAContext.h>
 
 #include "cutlass/cutlass.h"
 

csrc/quantization/w8a8/cutlass/c3x/scaled_mm_helper.hpp

@@ -1,4 +1,4 @@
-#include <torch/all.h>
+#include "../../../../vendors/functions.h"
 #include "cuda_utils.h"
 #include "cutlass_extensions/common.hpp"
 

csrc/quantization/w8a8/cutlass/c3x/scaled_mm_kernels.hpp

@@ -1,6 +1,6 @@
 #pragma once
 
-#include <torch/all.h>
+#include "../../../../vendors/functions.h"
 
 namespace vllm {