Fully transition from `flash-attn` to `kernels`

[qgallouedec]

The new recommended way to use flash attention is to use kernels. We should update our tests, and documentation to use kernels instead of "flash_attention2". Eg

trl/docs/source/reducing_memory_usage.md

Line 149 in 1eb561c

training_args = DPOConfig(..., padding_free=True, model_init_kwargs={"attn_implementation": "flash_attention_2"})

- training_args = DPOConfig(..., padding_free=True, model_init_kwargs={"attn_implementation": "flash_attention_2"}) 
+ training_args = DPOConfig(..., padding_free=True, model_init_kwargs={"attn_implementation": "kernels-community/flash-attn2"})