Added beyond-accuracy evaluation metrics

[giuspillo]

Hi @enoche,

thank you as always for this great job! I hope this can be interesting for you.

I've integrated the implementation for five beyond-accuracy metrics to the TopK evaluator, alongside the existing relevance-based ones (Recall, Precision, NDCG, MAP). Unlike those, these measure properties of how recommendations are distributed across the catalog, using item popularity computed from the training set.

• GiniIndex: inequality of the recommendation-frequency distribution (0 = uniform, 1 = one item gets everything).
• AveragePopularity: average training-set popularity of recommended items; higher = more popularity bias.
• TailPercentage: share of recommendations coming from the long tail (least-popular items), threshold configurable via tail_ratio. Default to 0.1.
• ItemCoverage: fraction of the catalog that appears in at least one user's top-K.
• ShannonEntropy: entropy of the recommendation-frequency distribution; complements GiniIndex as a diversity/equity measure.

Changes

• src/utils/metrics.py: implementations of the five metrics + item_metrics_dict registry, kept separate from the relevance-metric registry since they take item ids + popularity rather than a relevance mask. The four occourences of np.float have been replaced with np.float64 to adapt the current numpy versions.
• src/utils/topk_evaluator.py: TopKEvaluator now computes both metric groups and merges results into metric_dict following the order metrics are declared in the config (previously accuracy metrics always came first regardless of config order).
• src/configs/overall.yaml: all five item-distribution metrics added to the default metrics list; tail_ratio (default 0.1) added for TailPercentage as default value.