Recent update greatly increased GPU memory usage

[ducha-aiki]

Matching 2 images with 20000 matches works with commit 69cefb1
and doesn't with currentmain.

I guess that is because of removing

    torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul TODO: these probably ruin stuff, should be careful
    torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn

[Parskatt]

I think it's rather due to the implementation of the KDE which is naive: https://github.com/Parskatt/RoMa/blob/main/roma/utils/kde.py

20k matches from the KDE requires starting from 80k and then resampling down to 20k.

[Parskatt]

#22