Fix Windows CUDA build: guard extension_cuda compile options to CXX (#20184)

Guard extension_cuda's ${_common_compile_options} with $<COMPILE_LANGUAGE:CXX> so the MSVC /wd4996 flag no longer leaks (via slimtensor INTERFACE, added in #20158) into the aoti_cuda_shims .cu nvcc compile, which failed with 'nvcc fatal: A single input file is required'. Also run the cuda-windows workflow on extension/cuda changes. Verified: Windows CUDA e2e 5/6 green (was 0/6).

Author: shoumikhin

.github/workflows/cuda-windows.yml

@@ -16,6 +16,7 @@ on:
       - .github/workflows/cuda-windows.yml
       - backends/cuda/**
       - backends/aoti/**
+      - extension/cuda/**
   workflow_dispatch:
 
 concurrency:
@@ -49,6 +50,7 @@ jobs:
       (
         contains(needs.changed-files.outputs.changed-files, 'backends/cuda') ||
         contains(needs.changed-files.outputs.changed-files, 'backends/aoti') ||
+        contains(needs.changed-files.outputs.changed-files, 'extension/cuda') ||
         contains(needs.changed-files.outputs.changed-files, '.github/workflows/cuda-windows.yml') ||
         needs.run-decision.outputs.is-full-run == 'true'
       )
@@ -150,6 +152,7 @@ jobs:
       (
         contains(needs.changed-files.outputs.changed-files, 'backends/cuda') ||
         contains(needs.changed-files.outputs.changed-files, 'backends/aoti') ||
+        contains(needs.changed-files.outputs.changed-files, 'extension/cuda') ||
         contains(needs.changed-files.outputs.changed-files, '.github/workflows/cuda-windows.yml') ||
         needs.run-decision.outputs.is-full-run == 'true'
       )

extension/cuda/CMakeLists.txt

@@ -25,7 +25,9 @@ find_package(CUDAToolkit REQUIRED)
 add_library(extension_cuda SHARED caller_stream.cpp)
 target_link_libraries(extension_cuda PUBLIC CUDA::cudart)
 target_include_directories(extension_cuda PUBLIC ${_common_include_directories})
-target_compile_options(extension_cuda PUBLIC ${_common_compile_options})
+target_compile_options(
+  extension_cuda PUBLIC "$<$<COMPILE_LANGUAGE:CXX>:${_common_compile_options}>"
+)
 target_compile_definitions(
   extension_cuda PRIVATE EXECUTORCH_EXTENSION_CUDA_BUILDING
 )