Update to latest upstream LLaMA implementation

[philpax]

We're a couple weeks out of date with the current implementation of LLaMA in llama.cpp. There's quite a few changes (including always generating the BOS at the start!) that we should update to handle.

[AshleySchaeffer]

There's also this that's happened:

ggerganov/llama.cpp#1412

Which would help with GPU related issues I believe.