Configure how sequential and non-sequential features are combined

Session-based models are typically fed with sequential features (e.g. sequence of user interactions). But they can also be fed in addition with non-sequential features, that might represent:

• User/context features: e.g. user demographics (e.g. age, gender), user behavioural features (favorite item category, time since last purchase), context features (e.g. device, day of week).
• Target item features: e.g. when being used for ranking, you might want to provide features of the item target together with a label on whether the item is relevant or not the user/session. Those target item features are not sequential.

This task is about allowing users to configure via CLI how to combine sequential and non-sequential features:
Merge before the sequence processing

• Broadcast and concat - So that non-sequential features are replicated for the whole sequence and concatenated for each position

Merge after the sequence processing:

• concat
• element-wise ops: sum, mean, element-wise multiplication)

No files, tests, or entry points are named. Start by locating the CLI configuration and session-model sequence-processing paths, then trace how sequential and non-sequential features are currently passed through. Done means the CLI can select the listed pre-processing broadcast/concat and post-processing combination operations, with coverage for each supported choice.