[RMP] Enable users to pass embedding tables directly into the input block in order to more easily support new functionality (non-trainable embeddings, different dimensions, model parallel, etc)

[gabrielspmoreira]

Problem:

Customers want to be able to create embeddings that support a wide variety of options. The InputBlock handles the selection of these features through arguments, however we're up to 18 arguments and it's only going to get worse.
By separating input block definition from embedding creation we can allow for a much more flexible set of embeddings support with a simpler API.
The RMP #404 introduced the InputBlockV2 and also the Embeddings function that allow the user to provide a custom block for embeddings or infer the embeddings tables and their cardinality from the schema, exposing new options like setting trainable=False for embeddings.
But those new blocks are still experimental, for not being used yet by Merlin Models high-level API and have much less tests than the previous InputBlock

Goal:

• The Embeddings function is very handy to infer the embeddings definition from the schema. Make it more flexible by adding a trainable: Optional[List[str]] = None argument, to set EmbeddingTable.trainable=False (default is True) for the features present in the list. models#627
• [Task] Update DLRM to use the new EmbeddingTable models#613 - Update DLRM to use the new EmbeddingTable and InputBlockV2
• [Task] InputBlockV2 - Usage in ranking models high-level API models#604 - Update high-level API of ranking models - DCNModel, DeepFMModel - to use InputBlockV2 and to expose the embeddings block to the users.
• InputBlockV2 - Extend test coverage models#605 - Extend existing tests for InputBlockV2, by creating copies of existing tests for InputBlock, to ensure it is well tested - Moved this ticket to ENG-CI and closing the RMP ticket

Starting Point:

• The RMP [Task] Decouple InputBlock from Embedding creation allowing for user defined embeddings #404 is the starting point for introducing the InputBlockV2 and also the Embeddings function. The PR Supports task-specific sample weights, weighed_metrics and fixes InputBlockV2 models#600 introduced some fixes to InputBlockV2

Note:
non-trainable embeddings is the pre-trained embeddings

[viswa-nvidia]

Note comment for NVIDIA-Merlin/models#605