Memory leak in monai.transforms.distance_transform_edt

Describe the bug
There seems to be a memory leak in monai.transforms.distance_transform_edt when cuCIM is not available.

To Reproduce
cuCIM must not be available to reproduce

import torch
from monai.transforms import distance_transform_edt

nb_interations = 2000
for i in range(nb_interations):
    mask = torch.randint(low=0, high=2, size=(10, 10, 10,), device='cuda')
    distance_transform_edt(mask)

print(torch.cuda.memory_summary(device=None, abbreviated=True))

Expected behavior
The GPU memory footprint increases with nb_iterations. This should not happen. This does not happen when cuCIM is installed.

Additional context
The problem seems to be an exception leak in optional_import:

https://github.com/Project-MONAI/MONAI/blob/56508992c32e328667740e18f99b82a6a19f2c2a/monai/utils/module.py#L398-L408

where tb does not get cleared. tb holds references that prevent the mask variable from being deleted. I think this could be a problem with other optional_import calls, too.

Start with monai/utils/module.py around lines 398-408 and inspect optional_import, then reproduce the leak through monai.transforms.distance_transform_edt with cuCIM unavailable using the provided CUDA loop. Done means the repeated calls no longer cause GPU memory to grow, including for the reported fallback path.