[Task] Support pre-trained embeddings for ranking and session based models via the new dataloader functionality

[EvenOldridge]

• Update the input block to support the pre-trained embeddings:

• Add a module to extract pre-trained embeddings

• Show an example of how to transform pre-trained embeddings before aggregating them with the other trainable embeddings in the model. Such as Layer Normalization or MLP layer.

• Update DLRM, DeepFM, etc to include an optional projection layer for combining pretrained embeddings with the other embeddings in the model.

• Implement and evaluate different aggregation methods to combine pre-trained embeddings with the other trainable embeddings.

• Add an example demonstrating support for pretrained embeddings for DLRM or another non session based model. Artificial example? KDD cup?

Constraints

• Dataloader should support fixed-size 3D tensors.
• Dataloader needs to add a tag to the pre-trained embeddings so that the T4R input module can differentiate between trainable features and pre-trained embeddings.

[gabrielspmoreira]

Closing this macro-task, as the sub-tasks have been listed in this RMP