Bug found in supervision process

[HJoonKwon]

data['mask0'] and data['mask1'], indicating the padded regions, are coarse-level as presented in megadepth.py. while those masks are used to mask out the original-level matching points here. Even in case where those masks are not coarse-level, it's at least resized-level, not the original-level, as the mask was made after the image was resized in here. In both cases, the code has a bug in my opinion.