feat(inference): add vLLM inference

[maxreciprocate]

This PR adds vLLM inference in data parallel manner by launching multiple local servers in view of tensor_parallel_size

Usage:

from autocrit.inference import vLLMHook
# setup servers
model = vLLMHook("meta-llama/Llama-2-7b-hf", tensor_parallel_size=1, num_external_nodes=8)
model.generate(["..."], temperature=1, stop=["</s>"])
# stop servers
model.free()

Usage for reward models:

from autocrit.inference import RewardHook
model = RewardHook("reciprocate/rm_beluga-7b_hh-full")
scores = model.score(["..."])
model.free()