Allow different time resolutions from nwp datasource

[zakwatts]

Currently we use nwp datasets, where the temporal resolution of the forecast can change.

For instance, the Met Office UKV model, is at an hourly resolution for the steps 0-48 hours, then it goes to 3 hourly resolution for 48-54 hours.

Currently we do not support this changing of the resolution but, it would be a useful feature and allow us to train to longer time horizons more easily.

[Sagarpillai]

Heyy @zakwatts , I’ve been exploring this issue and would love to work on it.

A few things that I found out:

The loaders already pass through mixed step values correctly.

The problem seems to occur later in select_time_slice_nwp (ocf_data_sampler/select/time.py), where pd.date_range(..., freq=time_resolution) creates a uniform time grid which causes the misalignment when data switches from hourly to 3 hourly intervals.

Proposed approach:

1.	Replace the fixed-frequency pd.date_range with target times derived from the actual data — using combinations of init_time_utc + step.
2.	Filter these by the requested [start_dt, end_dt] window.
3.	Optionally thin (downsample) to time_resolution if needed, but without creating timestamps that don’t exist in the dataset.
4.	Keep dropout and selection logic unchanged.
5.	Add a small unit test using synthetic mixed-resolution data.

I’m new to opensource and OCF, but have been studying the codebase and would love to implement this if it sounds reasonable. Please let me know if I'm doing anything wrong, if not could you please assign this issue to me? :D