`ComputeWsiMetricsCallback` breaks for `.mrxs` images

[AjeyPaiK]

Describe the bug
When I use .mrxs images to train a segmentation model, ComputeWsiMetricsCallback breaks.
This is because when we use the writers on mrxs images, there is usually a slide bound for each WSI which offsets the image grid.

Recently, a PR was raised which added the feature to store the grid offset in the metadata of the cache. The image grid we make during validation to compute dice scores needs to be offset by this value. Then, it will be fixed.

To Reproduce
Train a segmentation model using a dataset containing .mrxs files. It should also have ROIs enabled.

Expected behavior
Training shouldn't break

Environment
dlup version: 0.6.1
How installed: pip
Python version: 3.11
Operating System: linux

Additional context
NA