Stream llama.cpp model download progress

[jbouder]

Summary

This changes the llamacpp capability so model downloads no longer happen silently inside llama-server startup.

Instead of launching bin/llama-server -hf ... directly from Pixi, serve now routes through a new Python launcher:

• pre-checks the requested Hugging Face GGUF repo and model selector
• selects matching model .gguf files plus any mmproj GGUFs
• runs hf download --dry-run to show cache/download intent
• streams hf download output to the terminal before server startup
• then starts llama-server -hf ... with the same CPU/GPU flags as before

The launcher keeps HF_HOME=models, so cached model files remain isolated to the capability and subsequent runs reuse the local cache.