feat: add experimental median_average_precision function

[shntnu]

Summary

• Add median_average_precision() as an experimental alternative to mean_average_precision()
• Uses median instead of mean for aggregating AP scores across grouped profiles

Motivation

We've observed cases where one or two outlier profiles within a group can disproportionately drag down the mean AP score, potentially masking otherwise strong phenotypic signals. Median-based aggregation may be more robust to these outliers.

Importantly, the null distribution calculation adjusts accordingly (using element-wise median of null distributions), so this should not produce optimistically biased results—the statistic is still tested in the same rigorous way as mean average precision.

This is an experimental function to explore whether median aggregation helps in these scenarios.

Details

The new function:

• Has identical signature to mean_average_precision() for easy drop-in usage
• Computes median of AP scores and normalized AP scores per group
• Uses element-wise median of null distributions for p-value calculation
• Applies the same FDR correction (Benjamini-Hochberg)

Test plan

• Added parametrized test for median_average_precision (with/without progress bar)
• Verified output columns exist and values are in valid ranges
• All 64 existing tests still pass

Next steps

• Add examples comparing mean vs median on datasets with outliers
• Request review after examples demonstrate the behavior

🤖 Generated with Claude Code

[shntnu]

Closing this PR based on experimental results: median aggregation doesn't improve retrieval overall (823 vs 875 significant groups, 25.0% vs 26.5%), with mixed results across annotation types.

Given the lack of clear benefit, I'd prefer to keep the API simple rather than add another aggregation option.

Details: https://github.com/broadinstitute/jump_production/issues/22#issuecomment-3764178955