support SAM finetune workflow for instance segmentation

[zzsi]

Implement an easy-to-use training pipeline for single-class instance segmentation.

SAM itself has limited object categorization capability. Text prompt was mentioned in the paper but not released in code.

Use grounded sam and [PerSam] (https://github.com/NielsRogge/Transformers-Tutorials/tree/master/PerSAM).

GLIP+SAM: https://www.reddit.com/r/MachineLearning/comments/12dx457/p_glip_sam_for_zeroshot_instance_segmentation/

Another example of using an object detector to prompt SAM https://github.com/RockeyCoss/Prompt-Segment-Anything

[zzsi]

Working on SEEM. Still need to work on a training pipeline for fine-tuning the model for instance segmentation.