Add vLLM integration

[rlouf]

In this PR we add an integration for the offline batching mode of vLLM with the following interface:

from outlines import models, generate
from vllm.sampling_params import SamplingParams


model = models.vllm("model_name", **vllm_specific_kwargs)
generator = generate.regex(model, "[0-9]{3}")

model.load_lora("path_to_lora")

params = SamplingParams()
answer = generator("Generate a number", max_tokens=10, sampling_params=params)

While we adopt the overall design of Outlines we let users specify vllm-specific parameters. They can also initialize the model using an instance of vllm.LLM.

TODO

• Test passing custom parameters to model init
• Test passing custom parameters to generation
• Handle sampling parameters
• Add load_lora method
• Add docstrings
• Add documentation
• Add issue for CFGLogitsProcessor with vLLM
• Open issues for refactor of llama.cpp integration