Update GGMLs to v2 of the model, which fixes some errors in training.

Initial GGML model commit. Requires llama.cpp from May 12th onwards.