Add flash_attention_1/flash_attention_2 support & examples

[truskovskiyk]

Are we supporting flash_attention feature? https://github.com/Dao-AILab/flash-attention/tree/main

[benjaminye]

This should be a relatively easy implementation. When loading a model with hf transformers, users can pass in attn_implementation="flash_attention_2". Details here

I suggest us:

1. freeze the version of transformers and trl in dependencies
2. add flash-attn into dependencies list
3. have ConfigModel mapping to parameters of the underlying transformers modules