Support GGJT v3

[philpax]

There is a new quantization level in llama.cpp, which means there will be models published with it in the near-future. We will need to support this.

ggerganov/llama.cpp#1508

[philpax]

To clarify: this has landed as v3 of the format with qnt level 2. We don't currently support this as we're trying to figure out if there's a good way to handle the breaking changes, but we're likely to update to it shortly.

[philpax]

This is now done, with the caveat that all of your v2 models will now be busted. #261 is the plan for that.