Update on "introduce shim layers for cudaguard and cudastreamguard"

### Summary

This diff introduces shim layers for CudaGuard and CudaStreamGuard in the Executorch project, which will be further used by cuda-AOTI models for stream/cuda control. The changes include:

*   Adding a new source file `runtime/shims/cuda_guard.cpp` and header file `runtime/shims/cuda_guard.h` to the `CMakeLists.txt` and `TARGETS` files.
*   Creating a new test target `aoti_torch_cuda_guard` in the `targets.bzl` file.
*   Defining the `cuda_guard.h` header file with the necessary includes, namespace definitions, and function declarations.

These changes aim to provide a shim layer for CudaGuard, which is responsible for handling CUDA-related functionality in the Executorch runtime. The shim layer will allow for better modularity and maintainability of the codebase.

Differential Revision: [D84126634](https://our.internmc.facebook.com/intern/diff/D84126634/)

[ghstack-poisoned]

Author: Gasoonjia

backends/aoti/CMakeLists.txt

@@ -44,8 +44,6 @@ target_link_options(aoti_common PUBLIC -Wl,--export-dynamic)
 target_link_libraries(
   aoti_common
   PUBLIC extension_tensor ${CMAKE_DL_LIBS}
-         # Link PyTorch libraries for AOTI functions
-         ${TORCH_LIBRARIES}
 )
 executorch_target_link_options_shared_lib(aoti_common)