Add NC validation metrics and fix mini-batch local updates

[Pulserius]

Summary

• Propagate validation splits through the NC data-loading and trainer paths.
• Report weighted global validation loss and accuracy during NC training, while keeping test evaluation final-only.
• Correct mini-batch NC training so each local step consumes one seed batch and performs one model update.
• Preserve compatibility with legacy NC data tuples that do not include validation fields.

Scope

Benchmark scripts, benchmark results, elastic-round experiments, and KubeRay/Prometheus work are intentionally excluded. They remain on update_hc_experiments.

Validation

• Compiled the changed NC modules with Python 3.12.
• Ran the targeted validation, data-path, trainer, mini-batch, and NC integration tests: 15 passed.
• The repository has unrelated pre-existing failures in broader unit suites; the equivalent federated-method failures are also present on main.