Speed up semantic segmentation annotations loading

[sokovninn]

This PR includes:

• Optimized computation for detecting unassigned labels to streamline automatic background addition.
• Improved performance of the SegmentationAnnotation.combine_to_numpy() function.

Loader benchmarks (1 worker, 5000 images):

• unassigned_pixels = np.sum(array, axis=0) == 0 - 3:15 (previous)
• unassigned_pixels = ~np.any(array, axis=0) - 1:58 (new)
• Without adding background - 1:35
• Optimized background and optimized SegmentationAnnotation.combine_to_numpy() - 1:45

Training benchmarks (1 epoch, 5000 images, 8 workers):

• Without optimization: 2:46
• With optimizations: 2:37
• Dummy labels: 0:56

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 92.26%. Comparing base (6bcbb11) to head (1a9ed09).
Report is 15 commits behind head on main.

✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #205      +/-   ##
==========================================
- Coverage   92.33%   92.26%   -0.08%     
==========================================
  Files          79       80       +1     
  Lines        4161     4265     +104     
==========================================
+ Hits         3842     3935      +93     
- Misses        319      330      +11     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.