export AOTI_TORCH_EXPORT on Windows. (#140030)

Fixes #139954

reproduce UT:
```cmd
pytest test/inductor/test_torchinductor_codegen_dynamic_shapes.py -k test_device_assert_dynamic_shapes_cpu
```
Issue:
<img width="856" alt="image" src="https://github.com/user-attachments/assets/5fc501a9-54e5-45ac-9fb3-509ec11a7abe">

After fixing:
![Image](https://github.com/user-attachments/assets/883846fb-8e92-4b9c-9400-daab32382a3a)

Reland:
1. Declare export on Windows explicitly.
2. Support cpu, cuda and xpu devices.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/140030
Approved by: https://github.com/jgong5, https://github.com/desertfire, https://github.com/malfet

Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>

CMakeLists.txt

@@ -1101,6 +1101,9 @@ if(NOT MSVC)
     append_cxx_flag_if_supported("-Wno-error=redundant-move" CMAKE_CXX_FLAGS)
   endif()
 else()
+  # Define export functions for AOTI.
+  add_compile_definitions(EXPORT_AOTI_FUNCTIONS)
+
   # skip unwanted includes from windows.h
   add_compile_definitions(WIN32_LEAN_AND_MEAN)
   # Windows SDK broke compatibility since version 25131, but introduced this

torch/csrc/inductor/aoti_torch/c/shim.h

@@ -44,8 +44,16 @@
 // to symbol clashes at link time if libtorch is included in a DLL and binary
 // that depends on the DLL. As a short term fix, we don't export the symbols.
 // In the long term, this will need to be addressed when Windows is supported.
-// #define AOTI_TORCH_EXPORT __declspec(dllexport)
+#ifdef OVRSOURCE
+// Do not export AOTI on Windows for internal builds
 #define AOTI_TORCH_EXPORT
+#else /* OVRSOURCE */
+#ifdef EXPORT_AOTI_FUNCTIONS
+#define AOTI_TORCH_EXPORT __declspec(dllexport)
+#else
+#define AOTI_TORCH_EXPORT __declspec(dllimport)
+#endif
+#endif /* OVRSOURCE */
 #else // !_WIN32
 #define AOTI_TORCH_EXPORT
 #endif // _WIN32