[Draft] Enable CUTLASS without host compiler

[LuFinch]

This is a draft PR to enable CUTLASS in torch-xpu-ops so that we can test cutlass kernels' accuracy/performance in Pytorch when SDPA/GEMM kernels are ready.

Since there is not determined plan of how to import cutlass-sycl repo, I download it in cmake for debug convinence.

I put all pytorch/aten wrapper functions in ATen/native/cutlass/*.cpp, which extract problem shape and device pointer from at::Tensor. It is compiled and link into libtorch_xpu_ops.a with pure gcc. They will call kernel launch functions from ATen/native/cutlass/sycl/*.cpp

I put all cutlass sycl kernels functions in ATen/native/cutlass/sycl/*.cpp. Since CUTLASS and syclcompat don't support -fsycl-host-compiler=g++, I compile the .cpp files of cutlass kernels into libcutlass_kernels.so library with pure icpx and then link it to libtorch_xpu_ops.a with gcc linker.

Currently, due to

1. libcutlass_kernels.so links to libtorch_xpu_ops.a
2. libtorch_xpu_ops.a links to libtorch_xpu.so
3. libtorch_xpu.so compiles/links aten/src/ATen/native/mkldnn/xpu/detail/Attention.cpp and aten/src/ATen/native/mkldnn/xpu/Attention.cpp
4. the include directory of torch-xpu-ops is exposed to libtorch_xpu.so.

Pytorch/aten/src/ATen/native/mkldnn/xpu/Attention.cpp can call wrap functions from torch-xpu-ops/src/ATen/native/cutlass/*.h directly like below

#include <ATen/native/cutlass/Attention.h>
at::native::cutlass_sycl::sdpa_backward(...);

I have verified that overrideable_sdpa_backward can call into YuanKun's cutlass sdpa backward kernel now. It passes a few accuracy UT.

[LuFinch]

@EikanWang Could you help take a look? I have verified that the aten .cpp in Pytorch can invocate functions from torch-xpu-ops's aten .cpp.

The issues need to discuss are

1. Where should we put the cutlass kernels? Pytorch or torch-xpu-ops?