Add 7B presets for Mistral

[tirthasheshpatel]

Closes #1418

Adds mistral_7b_en and mistral_instruct_7b_en presets for Mistral. Also added tests for presets for all the components of Mistral.

[mattdangerw]

Can you show a full generation run with the causal lm class running in a colab? Loaded from kaggle.

[tirthasheshpatel]

Can you show a full generation run with the causal lm class running in a colab? Loaded from kaggle.

Here: https://colab.research.google.com/drive/1508c8IY_nQQIsF33nUc_87l0mLo2BFw0?usp=sharing

[mattdangerw]

This looks great! Just made the Kaggle model public so we can re-run ci.

I think there is one small thing to fix here. With the Kaggle rewrite our tokenizers can be created without a vocabulary, which is loaded after the fact. Basically loading will look like MistralTokenizer.from_config(config); MistralTokenizer.load_assets(path). So we have to accommodate a tokenizer not knowing it's vocabulary on creation.

To deal with this fact, we only grab the special tokens from the tokenizer the first time a preprocessing layer built. Not when it is constructed. Basically we want this...

https://github.com/keras-team/keras-nlp/blob/bc3852f9fc3e9edaab7bc6c5d7ee40e6d3618ac2/keras_nlp/models/bert/bert_preprocessor.py#L141-L156

Instead of this...

https://github.com/keras-team/keras-nlp/blob/bc3852f9fc3e9edaab7bc6c5d7ee40e6d3618ac2/keras_nlp/models/mistral/mistral_preprocessor.py#L122-L132

You might be able to trigger a bug if you save a MistralCausalLM in the .keras format, and then load the model back, and check causal_lm.preprocessor.packer.start_value. Not sure but we should probably check.

[tirthasheshpatel]

Thanks for noticing this @mattdangerw! Should be fixed now. Let me know if it looks good to you now!

[mattdangerw]

lgtm!