Add loader for anemoi-inference projected grid (x/y) format

[Copilot]

A second anemoi-inference output format exists in the wild where the spatial domain is a structured 2D projected grid (y, x dims) rather than a flat values dimension, and the forecast reference time is stored as an explicit scalar data variable (forecast_reference_time) rather than being inferred from time[0].

New loader: anemoi_inference_xy

src/mlwp_data_loaders/loaders/anemoi/anemoi_inference_xy.py

Handles the format:

Dimensions:  (time: 1, y: 1069, x: 1069)
Coordinates:
  * time                     (time) datetime64[ns]
  * y                        (y) float32
  * x                        (x) float32
    latitude                 (y, x) float64
    longitude                (y, x) float64
Data variables:
    forecast_reference_time  datetime64[ns]
    projection               int32
    2t                       (time, y, x) float32
    ...

Key transformations in _preprocess (applied per-file before concatenation):

• Extracts forecast_reference_time → reference_time dimension coordinate
• Computes lead_time = time - forecast_reference_time and swaps it in for the time dimension
• Promotes latitude/longitude to coordinates if they arrive as data variables
• Drops forecast_reference_time and projection scalars

Output traits: forecast / grid / deterministic — consistent with the existing anemoi_inference loader.

Tests

tests/test_anemoi_inference_xy.py — 17 unit tests covering all _preprocess transformations and load_dataset behaviour (trait attrs, dimension structure, path coercion, variable preservation) using synthetic in-memory datasets.