Fix: enable multiprocessing-safe federated_fit with accuracy and data…

[SeanPeppers]

This patch improves the fashion_mnist_demo.py pipeline for local federated learning by:

• Adding last_accuracy to MyModule, satisfying the FLoX training API expectations
• Setting TORCH_DATASETS=./data fallback to avoid subprocess crashes during test evaluation
• Forcing spawn multiprocessing mode to avoid PyTorch fork() serialization issues

These changes make the example run cleanly and reproducibly on most systems.

Tested locally with 3 workers and 5 rounds using FedAvg.