Several small updates, as well as reverting the default training setup to automatically save out the best model according to validation metrics, rather than the last model.
- make sure validation loader always has at least one sample
- allow for early stopping again, as well as saving out model weights every n epochs
- make training fully deterministic
- increase flexibility of LitDaliWrapper to work with custom dali video pipeline functions
- increase flexibility of keypoint correspondence array for true multiview datasets