Bug Fix - Type error when evaluating metric depth on Kitti

[loevlie]

This pull request addresses the type error encountered when evaluating the fine-tuned Depth-Anything model on the Kitti dataset , a similar issue with ZoeDepth was reported here - #2113.

The error had two proposed solutions:

1. Downgrade the torch version.
2. Manually adjust the height and width in the model code.

I opted for the second solution, modifying the height and width manually. This decision was made to avoid potential compatibility issues with future versions of xFormers, which may require a more recent version of torch. Downgrading torch, while a viable short-term fix, could lead to more significant maintenance challenges down the line.

[NickLucche]

This also happens when running evaluate or depth_to_pointcloud more generally. Thanks for the quick fix!