[FEA] Adding feature-level attention

[marcromeyn]

🚀 Feature request

Adding feature-level attention to Merlin-models.

Motivation

In our retrieval experiments we had multiple instances where a simple MF-model was outperforming a Two-Tower model. We suspect this is because some of the features that were passed to the two-tower model were quite noisy. We would like to give the model a chance to learn to ignore these features. Our hope is that that would lead to two-tower models performing at least as good as MF.