feat: Improved training scripts and added updated pretrained weights

[frgfm]

This PR introduces the following modifications:

• deprecated LabelSmoothingCrossEntropy since PyTorch integrates it now
• revamped Mixup
• added wandb integration for experiment tracking
• updated training script by switching augmentation to TrivialAugmentWide, and selecting more appropriate optimizers
• updated pretrained weights of darknet19, cspdarknet53, cspdarknet53_mish, repvgg_a0, repvgg_a1, resnet50d
• updated classification perf report

[codecov]

Codecov Report

Merging #163 (e92d076) into master (cd3a624) will decrease coverage by 0.10%.
The diff coverage is 90.00%.

@@            Coverage Diff             @@
##           master     #163      +/-   ##
==========================================
- Coverage   93.97%   93.86%   -0.11%     
==========================================
  Files          51       51              
  Lines        3005     2985      -20     
==========================================
- Hits         2824     2802      -22     
- Misses        181      183       +2     

Impacted Files	Coverage Δ
holocron/models/darknetv2.py	100.00% <ø> (ø)
holocron/models/darknetv4.py	100.00% <ø> (ø)
holocron/models/repvgg.py	89.15% <ø> (ø)
holocron/models/resnet.py	94.82% <ø> (ø)
holocron/trainer/core.py	91.32% <89.28%> (-0.27%)	⬇️
holocron/nn/functional.py	92.25% <100.00%> (-0.19%)	⬇️
holocron/nn/modules/loss.py	90.90% <100.00%> (-3.00%)	⬇️