63. Debug: Loss Reduction

This cross-entropy loss function has a subtle bug in how it reduces the loss over the batch.

Signature: def buggy_cross_entropy(logits, targets)

• logits: (N, C) — unnormalized scores per class
• targets: (N,) — ground-truth class indices
• Returns: scalar loss

Contract: the returned scalar must be invariant to batch size for a fixed per-sample loss distribution — i.e. duplicating the batch should leave the loss unchanged. The implementation should match torch.nn.CrossEntropyLoss's default reduction.