CUDA: fix mul_mat_vec for CC 6.0

[JohannesGaessler]

Fixes #10318 (comment) .

The problem is that by default the code is being compiled for compute capabilities 5.2, 6.1, 7.0, and 7.5. A GP100 has compute capability 6.0, the minimum for FP16 intrinsics. The host code says that it can do MMV with those intrinsics but without GGML_CUDA_F16 there is no actual device code available. This PR is more of a band-aid fix that just makes GPUs with compute capability use FP32 arithmetic if the code was not compiled with GGML_CUDA_F16. Medium-term I intend to revise the handling of these intrinsics and I'll do a proper fix at that time.

[slaren]

Wouldn't be possible to check if the architecture is in __CUDA_ARCH_LIST__? Maybe we should do that more often instead of assuming that certain kernels are available.

[slaren]

Add to common code:

#ifdef __CUDA_ARCH_LIST__
constexpr bool ggml_cuda_has_arch_impl(int) {
    return false;
}

template<class ... Archs>
constexpr bool ggml_cuda_has_arch_impl(int arch, int first, Archs... rest) {
    return arch == first || ggml_cuda_has_arch_impl(arch, rest...);
}

constexpr bool ggml_cuda_has_arch(int arch) {
    return ggml_cuda_has_arch_impl(arch, __CUDA_ARCH_LIST__);
}
#else
constexpr bool ggml_cuda_has_arch(int) {
    return false;
}
#endif // __CUDA_ARCH_LIST__

Then:

Suggested change

	const enum ggml_prec prec = fast_fp16_available(cc) && cc != 600 ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
	const enum ggml_prec prec = fast_fp16_available(cc) && (cc != 600 || ggml_cuda_has_arch(600)) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;

This will keep the check at compile time, so it shouldn't add any overhead. Though I am sure you could come up with a more generic check.