[Task] Support of session-based model definition with masking techniques for TF

[sararb]

Problem:

The session-based blocks defined in #451 are implemented as standalone tasks and we need to link them together to provide known session-based models.

Goal:

• Implement the high-level blocks (inputs block, next item prediction task) needed for defining sequential models.

Constraints:

• [Task] First support of session-based recommendation with RNNs block #451 should be finalized before starting this work
• Update examples and unit tests with the new definition of InputBlock, Masking and Next Item Prediction task.

Starting Point:

• Simplify the InputBlock
  [Task] Decouple InputBlock from Embedding creation allowing for user defined embeddings #404
  Notes:

• Simplify the definition of input block by exposing the embedding argument to the user and remove the current nested structure to support 2-D and 3-D inputs.
  [Task] Decouple InputBlock from Embedding creation allowing for user defined embeddings #404

• Create a SequenceBasedInputBlock that accepts context and sparse schemas. Dropped as the new InputBlockV2 accepts multiple branches and also automatically broadcast context features to 3-D scores.

• Add to-2D and to-3D transformations as a post block in the InputBlockV2.
  -Add BroadcastToSequence Layer to broadcast context features to match sequence shapes.  models#737

• Support advanced masking tasks: CLM, MLM (PLM is excluded based on our research paper)
  Introduces Masked Language Modeling for Transformers models#780

• Implementation of Transformer block based in the configs defined in HuggingFace - transformers-based library.
  Note: This is an example of how Transformer-based architectures are implemented in HF (as keras layers).
  Transformer Block (for HuggingFace support) models#771

• Refactor base model to include a new design of masking/target generation for loss calculation and metrics.
  Introduces Masked Language Modeling for Transformers models#780

• Update weight-tying strategy based on the new refactor of the InputBlock (embeddings are independent tables and model context would be removed)
  Introducing Outputs-V2 models#715

• Support setting the masking task for fit and evaluate
  Introduces Masked Language Modeling for Transformers models#780

• Example of transformer session-based recommendation for training and deployment
  MM transformer training and serving example models#1045