Support new objects in vos_inference.py

Thanks for your wonderful work!
But, as you mentioned: **"currently it only supports VOS datasets where all objects to track already appear on frame 0 in each video"**, which makes lots of benchmarks unavailable for testing, e.g. YoutubeVOS. These benchmarks are indispensable for VOS tasks, and you've already reported on their performance in your paper.
Could you please release the complete code? That will help a lot, thanks!