[frazane]

Adds an optional ensemble block to forecaster run entries, so the inference step produces N members per initialisation. This PR stops at GRIB on disk; ensemble verification comes in a follow-up PR. Stacked on #286: ensembles rely on output_variables to keep storage manageable.

runs:
  - forecaster:
      checkpoint: https://...
      steps: 0/120/6
      ensemble:
        members: 10
        seed: 0   # member m runs with seed + m
      output_variables: [2t, 2d, 10u, 10v, sp, msl, tp, z]

anemoi-inference produces one member per process, so the prepare step writes one config_{NNN}.yaml per member, writing to grib/{NNN}/ with the file names of a deterministic run and tagged with number, numberOfForecastsInEnsemble and dataType=pf. One inference_execute job per init time runs the members on one node, inference_resources.gpu at a time (default min(N, 4), one GPU each), seeding torch with seed + m before anemoi-inference starts. Each member has its own log and .ok file under logs/inference_execute/{run_id}-{init_time}/, a retry only reruns the members that failed, and the rule's .ok means all members exist. Until ensemble verification lands, verification, plots and MEC use member 001 and log a warning. The ensemble block is part of the run-specific hash (run ids without it are unchanged) and is rejected on the forecaster of a temporal downscaler.

Tested with stormy-plateau (anemoi-inference 0.11.1), 2 inits × 4 members on 2 GPUs: the GRIB keys read back correctly with the COSMO definitions, members differ, and a member rerun with the same seed is bit-identical.

Changes: ForecasterConfig and the JSON schema, inference_prepare.py, inference_execute (new inference_execute_members.sh), run_specific_hash, a run_grib_dir helper used by the verification, plot and MEC rules, README, unit tests.

Notes

• eps goes in grib2_keys, not encoding: the COSMO definitions have no eps key, and anemoi-inference reads it from either to pick the ensemble template for accumulated fields.
• Existing inference_execute outputs will be rerun once after this change: Snakemake records each rule's shell template and params, and both changed (the rendered command of a deterministic run is identical). To keep existing outputs, remove their records once with snakemake --cleanup-metadata <files>. New runs are not affected. feat: select the variables written to the inference GRIB output #286 has the same effect through the new output_variables param of inference_prepare_forecaster.