Allow saving, loading and pushing adapter compositions together

[calpt]

Closes #441; closes #747.

This PR introduces a set of new methods for saving, loading and pushing entire adapter compositions with one command:

• save_adapter_setup()
• load_adapter_setup()
• push_adapter_setup_to_hub()

They require two main params:

• adapter_setup: the adapter composition to be saved. Identical to what can be specified for active_adapters
• head_setup: for models with heads, the head setup to save along with the adapters. Identical to what can be specified for active_head

Docs here

Example

model = AutoAdapterModel.from_pretrained("roberta-base")

# create a complex setup
model.add_adapter("a", config=SeqBnConfig())
model.add_adapter("b", config=SeqBnConfig())
model.add_adapter("c", config=SeqBnConfig())
model.add_adapter_fusion(["a", "b"])
model.add_classification_head("head_a")
model.add_classification_head("head_b")
adapter_setup = Stack(Fuse("a", "b"), "c")
head_setup = BatchSplit("head_a", "head_b", batch_sizes=[1, 1])
model.set_active_adapters(adapter_setup)
model.active_head = head_setup

# save
model.save_adapter_setup("checkpoint", adapter_setup, head_setup=head_setup)

# push
model.push_adapter_setup_to_hub("calpt/random_adapter_setup_test", adapter_setup, head_setup=head_setup)

# re-load
# model2 = AutoAdapterModel.from_pretrained("roberta-base")
# model2.load_adapter_setup("checkpoint", set_active=True)

[TimoImhof]

Looks good to me!