Add out of domain test datasets for benchmarking finetuned models

[anwai98]

The idea for this would be to have scripts unified in one place for evaluating automatic and interactive segmentation capabilities of micro-sam on a diverse set of open-source datasets (i.e. download them, preprocess them, run inference with new finetuned models and generate relevant metric scores) (ideally, we would like to support inference in both 2d and 3d - wherever relevant).

[anwai98]

This is resolved by #728.