Add type constraint to generation

[rlouf]

This PR allows to limit the generation to ints or floats. Closes #90.

The implementation with the OpenAI API is very limited since OpenAI doesn't allow to send more than 300 logit_bias values when ~900 tokens in their vocabulary correspond to ints. API:

import outlines.models as models

complete = models.text_completion.hf("sshleifer/tiny-gpt2", max_tokens=128)
question = "When I was 6 my sister was half my age. Now I’m 88 how old is my sister?"
answers = complete(question, type="int")

Also closes #96