Allow user to compile with any cuda version using github actions

[jianlins]

As the official releases have limited versions, this would fill the gap.

[ngxson]

IMO it's quite strange to include a workflow that is not intended to be used by us (but instead, being used by user via a fork). The workflow is never be triggered automatically, so maintainers never know if it passes or not. This defeats the whole point of having a CI.

A cleaner approach would be that user creates a blank repo containing this workflow, and the actions/checkout step clone source code directly from https://github.com/ggerganov/llama.cpp