use flash_attn_with_kvcache for faster inference

[vince62s]

This PR does:

1. switch apex RMSNorm to awq_inference_engine RMSNorm which is much faster
2. add rotary_theta as an option (llama/mistral used to use 1e4 while Mixtral and mistralv0.2 use 1e6)
3. use flash2 flash_attn_with_kvcache instead of regular flash_attn_func for step > 0 (much faster to increment cache in place in certain cases)