- Default img_size changed from 224 to 384 (already plumbed through
  build_transforms/build_dataloaders, just needed passing from main.py)
- WeightedRandomSampler now uses sqrt(1/count) instead of 1/count to
  reduce over-prediction of minority classes against visually similar
  majority class (Healthy vs Mild NPDR)
- Default batch size reduced 64->32 to accommodate larger images
- 4-epoch warmup-only test: macro F1 0.45, weighted F1 0.74, kappa 0.62
  already exceeding previous 22-epoch best on all metrics

- WeightedRandomSampler for balanced class sampling during training
- FocalLoss (gamma=2) for hard-example mining, no alpha (sampler handles balance)
- Stronger augmentation: RandomResizedCrop, ColorJitter, GaussianBlur, 30deg rotation
- Discriminative LR: backbone at 0.1x, head at 1x
- 3-epoch linear warmup + cosine decay scheduler
- Early stopping on macro F1 instead of weighted F1
- Updated defaults: batch_size=64, epochs=50, patience=10
- Comparison chart now shows both macro and weighted F1
- DataParallel checkpoint handling (strip module. prefix)
- Enable cudnn.benchmark for training speed