[FEA] Simplify the Transformer API

[sararb]

🚀 Feature request

The current API provides support for masking and inference through specialized Merlin Blocks, such as ReplaceMaskedEmbeddings and SequenceMaskLastInference. These blocks need to be explicitly set by the user in the pre/post arguments of the TransformerBlock. However, this adds complexity to the high-level MM API, as the user needs to be familiar with these custom blocks in order to correctly define and train a transformer-based model using CLM or MLM approaches.

It would be helpful to have a more streamlined approach for incorporating masking and inference support into transformer models, without requiring users to have extensive knowledge of specialized Merlin Blocks.