TeaCache for the Flow matching part to accelerate it 2x+

As the model uses Flow matching for decoding, it's theoretically possible to integrate TeaCache into it and get ~2x speedup.

TeaCache is opensourced and has a paper https://github.com/ali-vilab/TeaCache

Although, it its written about the timestep embedding there, the inputs in principle can be the bare input values as well

Currently, a generation of a picture takes 5 minutes on my single 4090, and it will much help to accelerate the process

Start by locating the Flow matching decoding path in the Bagel repository, then read the linked TeaCache implementation and paper. The issue names no files or tests, so first establish where caching inputs can be integrated and how generation speed is measured. Done means a working integration with a measured speedup for image generation.

Written by the indexing model from the issue text.