Support dataset with missing timesteps

[JPXKQX]

Is your feature request related to a problem? Please describe.

There are some datasets with missing timesteps and we might want to use them for training.

Describe the solution you'd like

These datasets contain a "missing" property containing the indices of the missing timesteps

from anemoi.datasets.data import open_dataset

ds = open_dataset("dataset.zarr")
ds.missing

>>> {4217, 3782, 492, 2868, 2073, 2842, 2843, 2844, 189}

One option is to compute the valid_indices based on the missing timesteps, the number of steps to roll out, the time increment, and the number of previous indices included as predictors.

Describe alternatives you've considered

No response

Additional context

No response

Organisation

ECMWF