[BUG] ImageStats crashes when precomputed 'nda_croppeds' is provided

[bluehyena]

Describe the bug
monai.auto3dseg.ImageStats.__call__() crashes when precomputed 'nda_croppeds' is provided in the input mapping.

The current implementation only defines the local 'nda_croppeds' variable when the key is not present, but later always uses 'nda_croppeds' to populate 'cropped_shape' and intensity statistics. As a result, passing precomputed cropped foregrounds raises 'UnboundLocalError'.

To Reproduce
Steps to reproduce the behavior:

import torch
from monai.auto3dseg.analyzer import ImageStats

analyzer = ImageStats(image_key="image")
analyzer(
    {
        "image": torch.arange(64.0, dtype=torch.float32).reshape(1, 4, 4, 4),
        "nda_croppeds": [torch.ones((2, 2, 2), dtype=torch.float32)],
    }
)

Expected behavior
UnboundLocalError local variable 'nda_croppeds' referenced before assignment.
If nda_croppeds is already provided, ImageStats should use it directly instead of recomputing foreground crops or crashing.

Screenshots

Environment
MONAI version: 1.5.2+87.g853f7023
MONAI rev id: 853f702
Python version: 3.9.12
Pytorch version: 2.8.0+cpu
Numpy version: 1.24.3
Platform: Windows-10-10.0.22631-SP0
Has CUDA: False

Additional context
While investigating this path, I also confirmed that the analyzer-side grad-state leak discussed in #5889 is still reproducible on the current dev branch when exceptions are raised before the restore line.

I have a local fix prepared with regression tests for both the nda_croppeds crash and the grad state leak.