log top-scoring samples to wandb (resolves #575)

[sammlapp]

This PR adds logging of highest n scoring clips per class to WandB after running model.predict()

• defaults to 3 clips per class for up to 5 classes
• moves wandb sample logging configuration to model.wandb_logging dictionary with 3 parameters: "n_preview_samples" (number of samples to log to wandb before training/prediction); "top_samples_classes": which classes to log top-scoring samples of after prediction (if None, first 5 classes or all if less than 5); "n_top_samples": how many top-scoring samples to log for each class

It also resolves a bug by setting self.has_clips=True in AudioSplittingDataset initialization. It should be True, but was False because of the way it was initialized and oversight when writing the code.