Implemented FlashAttention 2

I managed to get [FlashAttention-2](https://github.com/Dao-AILab/flash-attention) implemented by switching the LlamaAttention in models.llama.config with a flash attention monkeypatch I got from alpaca_lora_4bit and updated it to version 2 by using replacing unpadded (old v1) with varlen (v2). Here is [my fork](https://github.com/Jeduh/petals) with the changes.

Results:
- \>1.1x increase in inference RPS
- 2x increase in forward RPS

All though, I cant do much with this since the main swarm rejects this code change and sends a protocol unsupported error. Makes sense.

Could FlashAttention 2 please be implemented into petals? As I've demonstrated, It's really easy and it saves a huge amount of attention memory with the new version, as attention is now made linear as seq_len grows, allowing for much higher --attn_cache_tokens. It even works for Multi-query attention and grouped-query attention for Llama 2 and I feel like this is right in line with the projects intentions of batching and linear attention compute being the same cost for the 3000-4000th token as the 0-1000th making usage calculations easy. At least as an option for hosts to get more out of their system?