fix(windows): always define _WIN32 preprocessor macro to prevent PyTorch compiling unsupported code (#275)

PyTorch (especially dynamo) relies on the preprocessor macro `_WIN32` to
guard compilation of code sections in some places when targetting
Windows.

This macro is defined by MSVC defacto but, NVCC and others are not
setting it by default which can lead to compilation errors.

This PR ensures the macro is defined when targetting Windows platform so
we can safely rely its availability for the preprocessor.

Author: mfuntowicz

build2cmake/src/templates/cuda/preamble.cmake

@@ -104,6 +104,11 @@ message(STATUS "Rendered for platform {{ platform }}")
 {% if platform == 'windows' %}
 include(${CMAKE_CURRENT_LIST_DIR}/cmake/windows.cmake)
 
+# This preprocessor macro should be defined in building with MSVC but not for CUDA and co.
+# Also, if not using MVSC, this may not be set too ...
+# So we explicitly set it to avoid any side effect due to preprocessor-guards not being defined.
+add_compile_definitions(_WIN32>)
+
 # Generate standardized build name
 run_python(TORCH_VERSION "import torch; print(torch.__version__.split('+')[0])" "Failed to get Torch version")
 run_python(CXX11_ABI_VALUE "import torch; print('TRUE' if torch._C._GLIBCXX_USE_CXX11_ABI else 'FALSE')" "Failed to get CXX11 ABI")