Add Gemma checkpoint support

[rasbt]

Adds the new Gemma models by Google.

• Implement model download
• Test tokenizer
• Implement HF checkpoint conversion
  • 2B model
  • 7B model
• Make sure generate.py produces reasonable outputs
  • 2B model
  • 7B model
• Add chat template for the -it (aka instruct) versions
• Update download docs
• Update finetuning docs
• Test pretraining
• Test finetuning
  • Full finetuning
  • LoRA
  • Adapter
• Add unit tests
• Update README

Fixes #940

[rasbt]

1. The 2B model loads ok, the 7B doesn't yet. There seems to be a size mismatch.
2. The 2B model does produce garbage outputs in generate.py though; perhaps due to the missing layernorm?
3. I am using GeLU for now as GeGLU (the version I am using) also causes size issues, and then there's the fact that HF and Keras use GeLU according to our discussion in Support Gemma #940
4. Is the multi-query attention setting correct? In the conig.json, they set "num_key_value_heads" equal to the number of heads, which is n_query_groups=1 or n_query_groups=n_heads?

[rasbt]

Btw @carmocca or @Andrei-Aksionov please feel free to continue this PR if you have time and are interested. I may have to put in a stop for now due to another project that is due soon ... I thought it was more of a simpler port and thought that's something I could do in ~1 h but it appears I got a bit stuck here and have to put in a break for now.

[Andrei-Aksionov]

I cannot make changes to others PRs, so I guess Carlos is the only one :).

thought that's something I could do in ~1 h

So many of us have been in this situation before ...

[rasbt]

Thanks for the updates so far and getting the 2B to work!

[Andrei-Aksionov]

@rasbt Do we really need all of these steps?

- Update finetuning docs
- Test pretraining
- Test finetuning
   - Full finetuning
   - LoRA
   - Adapter

I did a sanity check with LoRA fine-tuning and had no errors. And there couldn't be any errors, since we didn't change anything related to training. Plus we have tests that check if LoRA and Adapter are appliable to the model.
Fine-tuning docs also don't care about a new model.

[carmocca]

Excellent job!