make CTD GenerativeSampler visibility-aware - #3305

Summary
A regression analysis revealed that performance for ctd_coam drops since rc12, and more specifically since #2995. This PR applied some improvements to the PyTorch DLCLoader, and amongs other changes it introduces a method _remove_nans() that sanitizes the ground truth keypoints.

Cause:
The CTD GenerativeSampler at deeplabcut/pose_estimation_pytorch/data/generative_sampling.py relied on NaN propagation as an implicit safety net: invisible/unlabeled GT joints carried (NaN, NaN, vis=0) and NaN-based math made every synth path (good/jitter/miss/inv/swap) produce vis=0, and a final nan_mask zero’d visibility. Net effect: invisible joints stayed invisible in the conditional input. Since this implicit assumption is now not met anymore, we should explicitly make the GenerativeSampler visibility aware.

Changes:
This PR introduces a simple fix, where for invisible keypoints the generative sampling logic is skipped and a visibility of 0 is always propagated:

for j in range(self.num_keypoints):
      if keypoints[j, 2] <= 0:
          synth_joints[j] = 0  # (x, y, vis) = (0, 0, 0)
          continue
          
       ....

Note that this fix restores the observed regression.