Fix FNO inverse pipeline for InitialConditionInterp with multi-step inputs

[Mattia-PCS]

Summary

This PR fixes the FNO inverse pipeline when using InitialConditionInterp with initial_step > 1.

Two related issues were identified:

1. The FNO branch did not initialize inverse_loader, preventing correct model initialization.
2. InitialConditionInterp reconstructed only a single initial time slice, while FNO checkpoints trained with initial_step > 1 expect the full initial temporal window.

Changes

• initialize inverse_loader in the FNO branch
• reconstruct the complete initial temporal window for FNO
• flatten the temporal window before calling the FNO model
• keep inverse metrics consistent with the flattened representation

Validation

The original code raises:

UnboundLocalError

After fixing the loader initialization, the next failure is:

RuntimeError: mat1 and mat2 shapes cannot be multiplied (64x4 and 16x20)

With this patch, the same command completes successfully:

loss=0.8711221814155579 ... 100%|...| 20/20

The resulting input matches the FNO forward pipeline:

u0 before flatten: (1, 64, 5, 3)
u0_for_model:      (1, 64, 15)
grid:              (1, 64, 1)
fc0.in_features:   16

Fixes #ISSUE_NUMBER