Automatically convert heads when loading with XAdapterModel

[calpt]

Related to #592.
Changes loading logic of XAdapterModel classes such that static heads of loaded checkpoint will be automatically converted.
E.g.

model = AutoAdapterModel.from_pretrained("bert-base-uncased")

print(model.heads)

should automatically create a masked-LM head using the weights of the original head from the pre-trained checkpoint.

[calpt]

There's currently still an open issue for models such as BERT, where the LM head bias is not correctly loaded:

from adapters import AutoAdapterModel

model, loading_info = AutoAdapterModel.from_pretrained("bert-base-uncased", output_loading_info=True)
print(model.heads)
print(loading_info)

The cause of this issue is unrelated to the current changes.

[lenglaender]

Looks solid to me 👍