* CUDA 11.0 fixes * Cleaner CUDA/host flags separation Also renamed GGML_ASSUME into GGML_CUDA_ASSUME