As discussed internally, we want to automatically cast the weights of
the adapter to float32 when using float16. Float16 is not conducive to
stable training and raises errors when used with AMP.

Previously, we had to recommend to users to manually cast the weights
if they loaded the base model in float16, because PEFT would choose the
same dtype for the adapter as for the base weights. Forgetting this is a
common source of errors, so we choose to automate this.

If this causes trouble, users can prevent the behavior by passing
autocast_adapter_dtype=False to get_peft_model,
PeftModel.from_pretrained, or PeftModel.load_adapter.

This PR should be reviewed carefully, as it has the potential to break
existing code if something important was missed. We also need to add a
note for the upcoming release text about this change in behavior.