make : add optional CUDA_NATIVE_ARCH (#2482)

Use the environment variable `CUDA_NATIVE_ARCH` if present to set NVCC arch. Otherwise, use `native`.

Makefile

@@ -391,9 +391,12 @@ else
 endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
 	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
+endif # CUDA_DOCKER_ARCH
+ifdef CUDA_NATIVE_ARCH
+	NVCCFLAGS += -arch=$(CUDA_NATIVE_ARCH)
 else
 	NVCCFLAGS += -arch=native
-endif # CUDA_DOCKER_ARCH
+endif # CUDA_NATIVE_ARCH
 ifdef LLAMA_CUDA_FORCE_DMMV
 	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV