Unify Offline Evals

[zentavious]

This PR introduces the iter_hyperparameters function to streamline hyperparameter tuning by generating and managing combinations of topic_theta and locality_theta values. It enhances the recommendation pipeline by supporting parallelized processing of profiles paired with these theta combinations, with an option for random sampling to reduce runtime. Additionally, it integrates key metrics for analysis and debugging.

New Functionality:

1. iter_hyperparameters: A generator function that yields cross-product combinations of topic_theta and locality_theta values paired with profiles.
2. Random sampling capability to reduce computational overhead (suggested by @mdekstrand).
3. Parallelization-ready structure for downstream components.

Pipeline Integration:

1. Added theta parameters to the rerank component.
2. Wired these parameters through the recommendation pipeline.

Metrics Logging:

1. KL Divergence: Measures the divergence between distributions.
2. Treatment Ratio: Tracks the ratio of profiles treated with specific configurations.
3. Within Locality Threshold Flag: Indicates compliance with locality-based constraints.