This pull requests makes some small updates and bug fixes:

Addresses #2703 to deal with NaNs when logging images to WandB

• The issue comes from NaNs being converted to integers
• With torchvision>=0.18, we can use a visibility flag. I think that's too restrictive for now, added a comment and we can pin torchvision in the future.
• Tests were added to ensure this works.

Addresses #2733 for modified project configuration filenames

• This is for consistency in the codebase. It is not regular DeepLabCut workflow to have a project configuration file named anything other than config.yaml, and no other method is guaranteed to work.

Fixes a bug when calling create_training_dataset with a non-default detector for multi-animal projects:

• detector_type=detector_type, was not passed to create_multianimaltraining_dataset

When calling video inference with cropping=True, updates the metadata for the predictions correctly:

• So the keypoints are displayed correctly when calling create_labeled_video and analysis is correct

PAF predictor - Fixes a bug when there are no detections

• The assembler returns None instead of an empty list when no assemblies are found
• This deals with it by checking that the assemblies aren't None

Heatmap target generation - Fixes a bug where gradients might be masked if there is missing data

• also adds gradient masking configuration to the configuration files