Model implementation details leak outside the model

Since the base Model class doesn't implement (at least a placeholder for) the call() method, other parts of the library can't rely on that method being defined. As a result, the details of how to compute predictions from a model show up in multiple other places:

• The example model evaluation
• The retrieval task
• The topk layers (1, 2)
• The topk metrics

There are two flavors represented among those examples:

• Those with one-to-one relationships between query embeddings and candidate embeddings (like the topK metrics) that are implemented with element-wise products
• Those with many-to-many relationships between query embeddings and candidate embeddings (like the retrieval task) that are implemented with matmuls

In order to consolidate the prediction code and abstract the way predictions are computed, it seems like these two modes could either be captured in a single Model method with a flag that selects between element-wise pairs and batch predictions, or represented as two Model methods.

Thoughts?