[ERMP] Make embedding layers more performant and scalable

[viswa-nvidia]

Problem

Customers need to be able to train models that exceed the memory of a single GPU. Recommender Systems models can grow to over hundreds of millions of users and items , which isn’t feasible to store on a single GPU.

Goal

• HugeCTR SOK implementation in Models library
• Make embedding layers more performant - ( Based on JoC example - 40% more efficient )

##Stories

• [Task] Decouple InputBlock from Embedding creation allowing for user defined embeddings #404
• [RMP] Model Parallel Embedding Support in merlin-models models#451
• Performance improvement of the embedding tables

[EvenOldridge]

Closing as a duplicate of #733