[DPOTrainer] Fix DPO trainer + mistral + FA2

[younesbelkada]

Fixes: #1217
Fixes: huggingface/transformers#26877
Fixes: #1266

Simply setting use_cache=False circumvents all issues with FA-2 + DPO + Mistral. in fact we should bypass that check since we are not in text generation mode when computing the loss function. use_cache is retrieved from the model config by default which falls back always to True. The cache is not used anyway when purely computing the logits so this change is fully BC

cc @kashif @vwxyzjn

To test that, I managed to repro the issue by adding --attn_implementation "flash_attention_2" in the dpo shell script on a A100 machine, and I confirm this PR fixes it. Unfortunately our CI runners are not compatible with FA2 so we cannot add a slow test to test that

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[kashif]

thank you great help!