Add LanguageModel Op and integration with HuggingFace's GPT2 implementation

[rlouf]

• Add a generic LanguageModel Op
• Transpile scripts containing language models
• Integrate HugginFace transformers
• Test on workflows that consist of several calls to a language model

[rlouf]

I plugged-in HuggingFace's GPT2 and I get the following results:

import outlines


gpt2 = outlines.text.models.GPT2()

output = outlines.script("""
Here is a good joke: ${joke}
And a random fact: ${fact}
""")(joke=gpt2, fact=gpt2)

fn = compile([], [output])
print(fn())
# Here is a good joke:  The last day of 2013  was quite an emotional and fun one.  I wrote,
# And a random fact:  I am going to be living here for a few years with my friend James.  

We still need to fix the following before merging:

• The LanguageModel instances are stateful: if I use the joke LanguageModel instance for joke and fact I get the same generation. This needs to be fixed.
• They are not random: running the program twice gives the same results.
• HF raises a warning when decoding without an input mask. We should figure out why and make sure it is not raised.
• Add tests for LanguageModel