fix: guard MSVC /utf-8 flag to CXX-only for engine_runtime

nvcc treats a bare /utf-8 as a second input file instead of forwarding
it to the host compiler, same failure mode already documented for
/openmp:experimental above. This only surfaced now because Release 0.5
added .cu sources (istft_cuda_runtime.cu, torch_random_cuda_runtime.cu)
to engine_runtime for the first time.

Author: owlab-developer

CMakeLists.txt

@@ -1187,7 +1187,10 @@ if (ENGINE_ENABLE_OPENMP)
 endif()
 
 if (MSVC)
-    target_compile_options(engine_runtime PRIVATE /utf-8)
+    # Guard on CXX for the same reason as /openmp:experimental above: nvcc
+    # treats a bare /utf-8 as a second input file instead of forwarding it
+    # to the host compiler, which is fatal for engine_runtime's CUDA sources.
+    target_compile_options(engine_runtime PRIVATE $<$<COMPILE_LANGUAGE:CXX>:/utf-8>)
 endif()
 
 if (ENGINE_ENABLE_CUDA AND NOT ENGINE_ENABLE_HIP)