Add support for probabilistic forecasts (and metrics)

[rodrigoalmeida94]

Add support for probabilistic (ensemble) forecasts and associated metrics throughout the codebase.

Key Considerations

• Performance & Memory Management:
  Ensemble-based metrics can be memory-intensive. The implementation should prioritize memory management to ensure scalability, especially for large ensembles. Efficiency optimisations can follow once memory usage is under control.

• API Design for Ensembles:
  The API should clearly define how ensemble forecasts are handled. Important questions include:
  What should the ensemble dimension be called (e.g., "member", "ensemble", "realization")?
  How should the API distinguish between deterministic and ensemble forecasts? Should this be inferred automatically, or should users be required to specify it?
  Should users be responsible for managing the ensemble dimension, or should the library provide utilities for parsing and handling both deterministic and ensemble inputs?

• Metrics:
  Implement and adapt probabilistic metrics (e.g., Brier Score, CRPS, ROCSS) and ensure they are performant and memory-efficient.

• Documentation:
  Update documentation to clarify how to use probabilistic forecasts, including API usage, expected input formats, and examples.

Open Questions:

• What should the default ensemble dimension name be?
• Should the API auto-detect ensemble vs deterministic, or require explicit user input?
• Are there existing libraries/utilities we can leverage for memory-efficient ensemble operations?

[linear]

EWB-90