Build a PyTorch DataLoader that balances class sampling with per-example weights derived from label frequencies.
