Use precomputed embeddings for SamAutomaticMaskGenerator

Is it possible to precompute the embeddings, similarly to predictor.set_image(), for the SamAutomaticMaskGenerator.generate() function?

This would be useful for quicker runs on a CPU machine after having precomputed the embeddings on a GPU.

Start with the SamAutomaticMaskGenerator.generate() and predictor.set_image() entry points named in the issue, comparing how image embeddings are prepared and consumed. Done means the generator can use embeddings precomputed on GPU for quicker CPU runs, with behavior matching normal generation.

Written by the indexing model from the issue text.