Model pipeline performance improvements

[deluzhao]

1. Handling in-line formulas and multiple bounding boxes corresponding to the same chunk/paragraph of text (e.g. "Hello" and "World" may be transcribed separately and we need to bring them together based on bounding box coords or using layout detection better)
2. Improving processing FPS (making specific models more lightweight, adding customization of model size and pipeline model selection) (e.g. remove Omniparser icon detection model and only use the underlying OCR model)