Add probabilistic evaluation metrics (CRPS, rank histograms) via existing scores dependency

[GiGiKoneti]

Hi team,

While going through the downstream validation pipeline for the neural-lam
probabilistic forecasting track (issue mllam/neural-lam#62), I noticed
that mllam-verification currently only covers deterministic
statistics like rmse and mae.

pyproject.toml already pulls scores>=1.2.0 which exposes exactly
what's needed for ensemble evaluation:

• scores.probability.crps_for_ensemble — for member-indexed ensemble
  outputs, which is the format neural-lam's datastore already uses
  consistently (ensemble_member dimension in weather_dataset.py
  and datastore/base.py)
• scores.plotdata.rank_histogram — for Talagrand diagram evaluation
  of ensemble calibartion

Two concrete additions that would follow the existing architecture exactly:

1. A crps() function in statistics.py wrapping
   scores.probability.crps_for_ensemble via compute_pipeline_statistic
   — same pattern as how rmse wraps scores.continuous.rmse
2. A plot_rank_histogram() in plot.py wrapping
   scores.plotdata.rank_histogram — following the plot_single_metric_timeseries
   structure

No new dependancies needed. Both functions already exist in the
pinned scores version.

If this makes sense I'll go ahead and implement it otherwise just
let me know and I'll close the issue.

[mfroelund]

Hi @GiGiKoneti

Thank you for your issue and the proposed addition to the mllam-verification package. The intention with including scores as a dependency, and the (hopefully) generic design of the compute_pipeline_statistic function is exactly to make it possible to add new metrics to the verification package. The current rmse and mae was just to have a limited set of metrics to get started from. So your proposed additions are very welcome, and I think it makes very good sense to get some metrics for evaluating ensembles given issue mllam/neural-lam#62.

So please go ahead and implement it:)

Best,

@mfroelund