fix(*:skip): Normalize avg_trainloss in PyTorch quickstart

[jun-simons]

Issue

Description

The PyTorch quickstart computes running_loss across all local epochs but normalizes it only by len(trainloader). As a result, the reported train loss metric scales approximately linearly with local-epochs > 1, which is misleading.

Related issues/PRs

Fixes #6333

Proposal

Explanation

This PR fixes the normalization by dividing running_loss by (epochs * len(trainloader)), yielding an average loss per batch over the full local training run.

This is a single line change to the train function, and the fix is also applied in the docs that use this code directly.

• examples/quickstart-pytorch/pytorchexample/task.py
• framework/docs/source/tutorial-quickstart-pytorch.rst
• framework/docs/source/tutorial-series-get-started-with-flower-pytorch.rst

Behavior for local-epochs=1 is unchanged but the train loss for local-epochs>1 is corrected. This was verified by running one server round with local-epochs={1,2,4} and the loss no longer scales 2x/4x, but instead remains in the same range.

Note that examples/quickstart-pennylane/... already normalizes loss as running_loss / (epochs * len(trainloader)).

Checklist

• Implement proposed change
• Write tests
• Update documentation
• Make CI checks pass
• Ping maintainers on Slack (channel #contributions)

Any other comments?

./framework/dev/format.sh currently fails on upstream main due to a unrelated missing copyright notice in framework/py/flwr/app/message_type.py, which is reproducible on main at 92b9de10b