Improve georeferencing (clipping) by using better keypoint matching method

> The failed georeferencing cases don't seem to be caused by the duplicated globes specifically — the globes aren’t always the ones producing the strongest or most ambiguous matches. The main issue seems to be that the current feature-matching pipeline (BRISK + FLANN) often produces incorrect matches between visually similar parts of the image, which then leads to a bad homography.
If we want to continue using the globes (or any repeated visual elements), one improvement could be switching to a more robust keypoint matching method. I actually tested SuperPoint + SuperGlue (https://github.com/magicleap/SuperGluePretrainedNetwork) with sketchmaptool last year — it worked very well and handled repeated patterns much more reliably.