Allow specifying own Hugging Face tokenizer instance

[shawnz]

This gives the option of specifying your own PreTrainedTokenizer instance rather than specifying a model name and having outlines construct the tokenizer for you. This might be convenient for situations where you are already using the LLM in other parts of your application and only need outlines for a specific use case.

Example of use:

from transformers import AutoTokenizer, AutoModelForCausalLM
from outlines.models.transformers import TransformerTokenizer, Transformer

model_name = "hf-internal-testing/tiny-random-GPTJForCausalLM"
hf_tokenizer = AutoTokenizer.from_pretrained(model_name)
hf_model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = TransformerTokenizer(hf_tokenizer)
model = Transformer(hf_model, tokenizer)

As compared to before this change:

from transformers import AutoTokenizer, AutoModelForCausalLM
from outlines.models.transformers import TransformerTokenizer, Transformer

model_name = "hf-internal-testing/tiny-random-GPTJForCausalLM"
hf_tokenizer = AutoTokenizer.from_pretrained(model_name)
hf_model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = TransformerTokenizer(model_name) # Can't make use of hf_tokenizer here, even if
                                             # you already instantiated it for other reasons!
model = Transformer(hf_model, tokenizer)

This also corrects a type annotation on the outlines.models.transformers.Transformer class which incorrectly stated that the tokenizer argument was of type transformers.PreTrainedTokenizer when actually it should be of type outlines.models.transformers.TransformerTokenizer. In order to fix the annotation, the classes also had to be rearranged so that the name of the other class would be defined at the right time.

[rlouf]

This part of the interface has been bugging me for a while, thank you for helping!