Report when GPU with an incompatible CUDA architecture is used

### What is the issue? When building Ollama, support for different CUDA capabilities can be toggled using the `-DCMAKE_CUDA_ARCHITECTURES` flag. When a GPU with an architecture disabled this way is used, it results in the following cryptic error message on the server followed by long stacktraces: ``` [GIN] 2025/07/02 - 14:20:13 | 200 | 6.234077207s | 127.0.0.1 | POST "/api/generate" ggml_cuda_compute_forward: RMS_NORM failed CUDA error: no kernel image is available for execution on the device current device: 0, in function ggml_cuda_compute_forward at /build/source/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu:2366 err /build/source/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu:76: CUDA error SIGSEGV: segmentation violation PC=0x7f7b16424c57 m=4 sigcode=1 addr=0x206203fe0 signal arrived during cgo execution ``` `ollama serve` already knows which compute capability does used GPU have, because it prints this information to stdout (the `compute=6.1` part): ``` time=2025-07-02T13:38:19.391+02:00 level=INFO source=types.go:130 msg="inference compute" id=GPU-8020c948-dcac-7cc5-4991-07408ef9edad library=cuda variant=v12 compute=6.1 driver=12.4 name="NVIDIA GeForce GTX 1060 with Max-Q Design" total="5.9 GiB" available="5.9 GiB" ``` I suggest reporting a warning when server detects that GPU's CUDA architecture is unsupported, so that the user knows that the issue is caused by incompatibility of GPU and ollama build. ### Relevant log output ```shell ``` ### OS Linux ### GPU Nvidia ### CPU Intel ### Ollama version 0.9.3