ci : faster CUDA toolkit installation method and use ccache

[slaren]

ccache could be enabled for other builds, it makes a big difference. When building PRs, it should reuse the cache from the parent build in master.

Credit where it's due: I took the CUDA installation method from the CI in karpathy llm.c

Fixes #10486

[JohannesGaessler]

If ccache actually helps we obviously should use it though I'm confused that it does. That would imply that there is persistent storage between different CI builds which to my knowledge there shouldn't be.

[slaren]

Github has a cache that allows data to be kept between builds. The name of the step is a bit misleading, it doesn't just install ccache, it also restores the ccache cache from the github cache.
https://github.com/ggerganov/llama.cpp/actions/caches