feat: use FindCUDAToolkit for robust CUDA detection

Author: SyntX34

CMakeLists.txt

@@ -20,16 +20,23 @@ if(USE_GPU AND NOT FORCE_CPU)
         set(GGML_METAL ON CACHE BOOL "" FORCE)
         set(SUBGEN_GPU_BACKEND "Metal")
     elseif(WIN32 OR (UNIX AND NOT APPLE))
-        include(CheckLanguage)
-        check_language(CUDA)
-        if(CMAKE_CUDA_COMPILER)
-            enable_language(CUDA)
+        find_package(CUDAToolkit QUIET)
+        if(CUDAToolkit_FOUND)
             set(GGML_CUDA ON CACHE BOOL "" FORCE)
             set(SUBGEN_GPU_BACKEND "CUDA")
+            message(STATUS "CUDA toolkit found -- building with CUDA support")
         else()
-            message(STATUS "No CUDA toolkit found at configure time -- building CPU-only. "
-                            "(GPU is still used automatically at runtime if you later "
-                            "rebuild on a machine with the CUDA toolkit installed.)")
+            include(CheckLanguage)
+            check_language(CUDA)
+            if(CMAKE_CUDA_COMPILER)
+                enable_language(CUDA)
+                set(GGML_CUDA ON CACHE BOOL "" FORCE)
+                set(SUBGEN_GPU_BACKEND "CUDA")
+                message(STATUS "CUDA toolkit found (fallback) -- building with CUDA support")
+            else()
+                message(STATUS "No CUDA toolkit found -- building CPU-only. "
+                                "(Install CUDA toolkit from nvidia.com and rebuild to enable GPU acceleration.)")
+            endif()
         endif()
     endif()
 else()