RL training of Llama 3.1 8b on multi-gpus

[AlexPiche]

Training Llama 3.1 8b on multi-gpus

The goal of this PR is to train Llama 3.1 70b on GSM8k in an efficient manner.

• Launch accelerate job as a subprocess. better multi-gpu inference and training in the RL example #76
• Fix the NCCL timeout when training on multiple gpus.
• Implicit KL like Eq3 https://arxiv.org/abs/2402.14740
• Optimize get log probs OOM when getting logprobs from vLLM reference model #77

[rizar]

will address #77 and #76