Gemma2 text model support

[spawner1145]

#321

models in https://huggingface.co/neta-art/gemma2-2b-gguf

[city96]

So I don't think we can merge this as-is, in part because it has a lot of random changes that look like AI slop to me (i.e. the logging messages with the checkmark emoji lol), but mainly because I don't think making custom model formats for text models supported by llama.cpp is acceptable.

We should only ever convert image models, and for the text models the conversion should be handled by mainline llama.cpp.

The tokenizer can be reconstructed from the metadata, though you need to do a tensor operation to actually make it usable, otherwise the norms will be messed up and you'll get bad results. I tried my best to try and implement it, though I couldn't get the tokenizer to be 1:1 (everything else is tested working with proper quants from huggingface) #358

If you have the time, you could try and see if you can find the issue with the above PR, I'm pretty sure it's either due to spm.trainer_spec or due to the individual piece.score values. I left the test line in to load the quantized model but replace sd["spiece_model"] to test as the baseline. I can go into more detail, if you have any questions.

(Also, this PR would break UMT5 support, which already uses the tokenizer reconstruction logic)