fix: Fixed multiple detection training & model issues

[frgfm]

This PR introduces the following modifications:

• refactored loss weighting of YOLO
• fixed loss weight, anchor scaling & head initialization of YOLOv2
• fixed classification loss and added contribution ignoring to YOLOv4
• adds feature extraction possibility to DarknetV3
• adds nan tolerance to trainer
• refactored metric printing for detection trainer
• fixed torchvision compatibility of detection training

[codecov]

Codecov Report

Merging #182 (3eae9bd) into master (10106c3) will decrease coverage by 0.33%.
The diff coverage is 75.51%.

@@            Coverage Diff             @@
##           master     #182      +/-   ##
==========================================
- Coverage   93.83%   93.49%   -0.34%     
==========================================
  Files          54       54              
  Lines        3019     3046      +27     
==========================================
+ Hits         2833     2848      +15     
- Misses        186      198      +12     

Impacted Files	Coverage Δ
holocron/models/darknetv3.py	84.74% <35.71%> (-15.26%)	⬇️
holocron/trainer/core.py	89.26% <50.00%> (-1.19%)	⬇️
holocron/models/detection/yolo.py	98.07% <100.00%> (+0.02%)	⬆️
holocron/models/detection/yolov2.py	98.83% <100.00%> (-0.03%)	⬇️
holocron/models/detection/yolov4.py	97.67% <100.00%> (+0.05%)	⬆️
holocron/trainer/detection.py	88.60% <100.00%> (+0.44%)	⬆️

[frgfm]

cf. #110