Fix MLM loss masking in PyTorch BERT pretraining

[Chessing234]

Summary

• nn.CrossEntropyLoss() defaults to reduction='mean', returning a scalar
• Multiplying a scalar by mlm_weights_X.reshape(-1, 1) makes the padding mask a no-op — every position contributes equally regardless of whether it was a real prediction or padding
• Changing to reduction='none' returns per-element losses so the weight mask correctly zeros out padded positions before the manual sum/normalize
• The NSP loss now calls .mean() explicitly since reduction='none' applies to the shared loss instance
• The MXNet version is already correct because gluon.loss.SoftmaxCELoss() returns per-element losses by default

Fixes #2582

Changes

• chapter_natural-language-processing-pretraining/bert-pretraining.md:
  • Line 58: nn.CrossEntropyLoss() → nn.CrossEntropyLoss(reduction='none')
  • Line 118: loss(nsp_Y_hat, nsp_y) → loss(nsp_Y_hat, nsp_y).mean()

Test plan

• Run the PyTorch BERT pretraining notebook end-to-end
• Verify MLM loss values differ from before (mask now has effect)
• Verify NSP loss values are unchanged (.mean() matches prior default behavior)

🤖 Generated with Claude Code