Add ctd integration test - #3040

This pull request adds a CTD model to the integration-tested models in examples/testscript_pytorch_multi_animal.py .

To pass the test successfully, a few fixes have been done to the CTD pipeline:

• When loading conditions from an existing .pickle file, the _assemblies.pickle file is read (instead of the _full.pickle)
• The CondFromFile.load_conditions_pickle() method is fixed (was not picking the conditions for the last frame)
• A fix is added to the FilterLowConfidencePoses Preprocessor stage (introduced in Improve CTD preprocessors pipeline #2984), which was failing when the BU was making no keypoints predictions for a frame.
• The create_video_with_all_detections() method is fixed (previously was assuming that if predictions contained a "bboxes" key they also contained a "bbox_scores" key - this is not the case with CTD models. Similar case fixed in Fix deeplabcut.analyze_images() with CTD model #2990).
• The PadOutputs Postprocessor stage is made more robust:
  • All outputs are now normalized to np.ndarray (previously, inputs that didn't need padding were not converted to np.ndarray)
  • The PadOutputs now takes an additional expected_shapes argument at __init__(), which it uses to pad correctly empty input arrays.