[2502][evaluation] score bug fixes

[moritzhauschulz]

Description

Enables the previously dead probabilistic metrics (spread, ssr, crps, rank_histogram) in the evaluation package, corrects the spread-skill ratio to the standard ensemble-mean definition, and makes the spatial score-map path robust for metrics that collapse the ensemble dimension. This unblocks GenCast-style spread-skill diagnostics over lead time and ensemble-spread maps.

I am not very familiar with the eval pipeline, so it would be good if this could be reviewed by someone more knowledgeable

scores/score.py

• Enable probabilistic-metric dispatch. Replace the dead assert self.ens_dim … / return None
  (undefined self.ens_dim, unconditional return) with a real check: warn and skip when the
  ensemble dim self._ens_dim is absent from the predictions (e.g. deterministic runs), otherwise
  dispatch to the metric function. This activates spread, ssr, crps, and rank_histogram.
• Correct the spread-skill ratio. calc_ssr now divides the ensemble spread by the RMSE of
  the ensemble mean (the "skill", GenCast / WeatherBench2 convention) —
  calc_spread(p) / calc_rmse(p.mean("ens"), gt) — instead of the full-ensemble per-member RMSE.
  SSR is now a single value per variable/level/lead-time with the standard calibration
  interpretation (under-/over-dispersion), consistent with the already ensemble-reduced spread
  numerator.

plotting/plot_orchestration.py (_plot_score_maps_per_stream)

• Fix CoordinateValidationError crash. Guard the ensemble-label assignment on
  "ens" in plot_metrics.dims (the concatenated result) rather than preds.dims. When every
  selected metric reduces the ensemble dim (e.g. metrics: ["ssr"]), plot_metrics has no ens
  dim and the previous unconditional plot_metrics["ens"] = preds.ens raised.
• Avoid redundant per-member maps. Track which metrics retain a per-member ens dim in their
  own result (ens_metrics). Iterate ensemble members only for those; ensemble-reduced metrics
  (spread, crps, ssr) get a single map instead of one identical map per member.
• Collapse broadcast ensemble axis. When xr.concat broadcasts a reduced metric across ens,
  select a single member (isel(ens=0, drop=True)) so the plotted field is 2-D.

Issue Number

Closes #2502

Checklist before asking for review

• I have performed a self-review of my code
• My changes comply with basic sanity checks:
  • I have fixed formatting issues with ./scripts/actions.sh lint
  • I have run unit tests with ./scripts/actions.sh unit-test
  • I have documented my code and I have updated the docstrings.
  • I have added unit tests, if relevant
• [] I have tried my changes with data and code:
  • I have run the integration tests with ./scripts/actions.sh integration-test
  • (bigger changes) I have run a full training and I have written in the comment the run_id(s): launch-slurm.py --time 60
  • (bigger changes and experiments) I have shared a hegdedoc in the github issue with all the configurations and runs for this experiments
• I have informed and aligned with people impacted by my change:
  • for config changes: the MatterMost channels and/or a design doc
  • for changes of dependencies: the MatterMost software development channel

[clessig]

@jpolz : could you have a look?