feat: llama.cpp gRPC C++ backend

[mudler]

Description

This PR fixes #1154, fixes #1017 and fixes #944

It provides a C++ gRPC server to use in place of the go-llama binding. note it is much more flexible as now the Makefile controls directly the commit/branch/tag of llama.cpp. There are of course lots of optimizations that can be done, but this provides the first steps.

There are still some TODOs some required for a feature parity with the golang backend:

• Expose lora parameters
• Embeddings (feat parity)
• Speculative sampling (feat parity)
• Parallel sampling
• Additionally expose things like logits bias and an infill endpoint

What supposedly should work:

Everything, except loras. Including also CUDA/etc support

Next:

• Merge it
• Merge and unblock ci: GPU tests #1116 so we test with GPUs automatically too
• expose lora (now it needs also a scale factor)
• move golang backends inside backends (nearby the cpp one)
• move extra inside backends
• make it feature-parity with the current llama backend
• sunset the current llama in favor of this one
• rename llama-stable to llama-ggml to indicate support for old ggml file format (prior to gguf)

[mudler]

reminder to myself: create issues for follow-ups that are not tackled here