Add Alpaca support to talk-llama example

[ejones]

I made a couple edits locally to get talk-llama working with alpaca-13b, and thought I'd contribute back if that's helpful.

The invocation I used is:

./talk-llama -mw ./models/ggml-base.en.bin -ml path/to/ggml-alpaca-13b-q4.bin -p Evan -t 8 --n-parts-llama 1 --verbose-prompt --prompt-file examples/talk-llama/prompts/talk-alpaca.txt

This basically boils down to making the prompt and n_parts configurable in the CLI as well as adding a sample prompt in the Alpaca format. Seems to work well with alpaca-13b; I don't have any other models to test it with. Also I admit that making the prompt configurable here exposes the {N}-style interpolation that talk-llama.cpp does, which may warrant more thought.

[JKeddo95]

Have you seen the llama_model_load: invalid model file '../llama.cpp/models/Alpaca/13B/ggml-alpaca-13B-q4.bin' (bad magic) message? I am able to run fine in llama.cpp, but can't get it started from whisper. Can you confirm you are running newest Alpaca.cpp? Which converter (if any) did you use for your models before running?

[jmtatsch]

I would very much like to talk to my vicuna but am suffering from the same bad magic error.
Is the ggml format maybe too new for the llama.cpp contained in whisper.cpp?