[FEA] Train a model with multiple optimizers

[sararb]

🚀 Feature request

Train different parts of the model with different optimizers.

Motivation

The feature request is mainly motivated by the implementation of the Wide&Deep model that jointly trains the wide and deep components with two different optimization algorithms.

Your contribution

TensorFlow Recommenders supports the CompositeOptimizer class that allows different optimizers to be applied to different subsets of the model's variables. Following the same structure, we can define a MultiBlockOptimizers class that allows different optimizers to be applied to different subsets of the model’s blocks.