refactor TransformerBlock for serialization

[sararb]

Goals ⚽

• Improve the signature HF transformers call to fix the issue of shape mismatch when saving the model. This error is linked to non-used arguments being initialized with wrong tensors: The previous call method provided transformer_kwargs twice, e.g. mask_perm was set equal to input_embed instead of None).

• Extract from the HF model the serializable layer TF*MainLayer. In fact, all the HF models are defined with one custom Keras layer containing the architecture graph. (e.g. tfXLNetModel)

Testing Details 🔍

• Update test_transformer.py to include Albert config.
• Add a unit test that checks the serialization of the TransformerBlock