Fix fp16 by converting outputs back to FP32

[sgugger]

As was pointed out in #101, there is currently a bug in mixed precision training: the outputs are properly computed in mixed precision but are returned in FP16, and the loss computation is not inside a torch.cuda.amp.autocast context manager so is executed in full FP16, which is generally unstable (especially for softmax).

This was not discovered with Transformers models as the loss is computed inside the model, which is generally a better idea if one wants to use mixed precision with Accelerate.

To fix the problem, this PR:

• automatically converts the outputs of the model to FP32 so that the loss is executed in full precision (generally a better idea)
• adds a context manager accelerator.autocast for more complex loss functions that should be executed in mixed precision (as with all things accelerate, this context manager always work, it just does nothing if FP16 is not activated).

[LysandreJik]

Looks great! Cool workaround on autocast. LGTM!