[PyTorch] Use reduction=none for train_epoch_ch3

[AnirudhDagar]

Description of changes:
Fixes loss not showing up in the following sections:

• https://d2l.ai/chapter_linear-networks/softmax-regression-concise.html
• https://d2l.ai/chapter_multilayer-perceptrons/mlp-scratch.html
• https://d2l.ai/chapter_multilayer-perceptrons/mlp-concise.html
• https://d2l.ai/chapter_multilayer-perceptrons/dropout.html

I'll look at other sections that might be affected. Note that l.mean() is required otherwise when using l.sum() with PyTorch CrossEntropy (with no reduction) the loss overshoots. This change will not affect any other notebooks because d2l.train_ch3 is only used in these notebooks, so it won't break any other section.

By submitting this pull request, I confirm that you can use, modify,
copy, and redistribute this contribution, under the terms of your
choice.

[mli]

Job d2l-en/PR-2007/1 is complete.
Check the results at http://preview.d2l.ai/d2l-en/PR-2007/