CUDA: fixed LLAMA_FAST compilation option

[JohannesGaessler]

The LLAMA_FAST compilation option which replaces -O3 with -Ofast currently does not work in combination with CUDA because nvcc does not have that compilation option. This PR makes it so that -O3 is always used for nvcc and -Ofast otherwise.

[slaren]

Looks good, but I wonder if LLAMA_FAST should toggle -use_fast_math for nvcc instead.

[JohannesGaessler]

All llama.cpp CUDA calculations at some point use half precision or worse so I don't think there's a point in disabling --use_fast_math. I don't know if there are side effects for using -Ofast for llama.cpp/ggml in general so I'm being conservative in not making it the default.

[ggerganov]

All llama.cpp CUDA calculations at some point use half precision or worse so I don't think there's a point in disabling --use_fast_math. I don't know if there are side effects for using -Ofast for llama.cpp/ggml in general so I'm being conservative in not making it the default.

At some point there was discussion about this in whisper.cpp and I decided to keep O3 as default:

ggml-org/whisper.cpp#252 (comment)