🚀[FEA]: Add InterpCRPSDiT for CRPS temporal interpolation

[gertln]

Is this a new feature, an improvement, or a change to existing functionality?

New Feature

How would you describe the priority of this feature request

Medium

Please provide a clear description of problem you would like to solve.

Problem

Earth2Studio's global forecast and reanalysis models operate at coarse (typically 6-hourly) time steps, but many downstream applications need hourly or sub-hourly fields:

• tropical-cyclone tracking / rapid-intensification, where a 6 h step badly under-samples the central-pressure minimum,
• renewable-energy and demand modelling,
• event-timing analyses.

There is currently no way to turn a coarse trajectory into a 10–15 minute stream with a CRPS-trained ensemble spread. InterpModAFNO provides deterministic temporal interpolation but does not produce an uncertainty-aware ensemble and is oriented at hourly output, so mid-gap uncertainty and sub-hourly cadence are not addressed.

Proposed solution

InterpCRPSDiT a one-shot, endpoint-pinned CRPS temporal-interpolation model (DiT backbone with neighborhood attention) that refines a base model's coarse trajectory to higher temporal resolution:

• splits each coarse bracket into num_interp_steps equal sub-steps (=6 --> 1 h, =24 --> 15 min, =36 --> 10 min); supports any whole-minute cadence that evenly partitions the coarse gap (num_interp_steps must divide the gap in minutes),
• pins the endpoints to the base model's own frames (out(tau) reduces to the base states at tau=0,1), learning only the interior correction,
• produces an ensemble whose spread is trained with a CRPS objective,
• runs on compatible 0.25° bases supplying the required variables, including bases with fewer channels (e.g. Pangu) via a drop_variables path and supports regional sub-domain inference via set_domain.

Implemented in PR #987 (stacked on the DataReplay PR #985 ).