modify encode_video() / encode_audio() API design

[awkrail]

This PR is on WIP. Modifying encode_video and encode_audio API design because the current code sets video_feats/mask to the Predictor, which is buggy in gradio demo. We make the models stateless, which makes us more easier to use.

[awkrail]

@h-munakata Could you review PR?

[h-munakata]

I found an unexpected behavior.
When we change the model and the model extract the video feature showing Processing the video. Wait for a minute..., pushing the retrieve moments button leads AttributeError: 'XXPredictor' object has no attribute 'inputs'.
I think this is caused model object does not have inputs before running model.encode_video()
Therefore, error handling is required when the button is pressed immediately after the model is loaded and model do not hold inputs.

In addition, I found a line that should be fixed [Link].(https://github.com/awkrail/lighthouse/blob/5b5d14e5394e5615b63b7e8e3d5820fa3f7da757/lighthouse/common/tr_detr_transformer.py#L141)
This line requires CUDA, and thereby, the inference does not work without GPU.