feat(validate): add precision, recall, and F1 metrics

[ha405]

Overview

This PR adds optional support for calculating Precision, Recall, and F1-score in validate.py.
These metrics provide a more nuanced view of model performance than Top-k accuracy, especially for datasets with class imbalance.

Changes

• New flag: --metrics-avg to select averaging method (weighted, macro, micro).
• Metric computation: Uses scikit-learn for robust precision/recall/F1 calculations (soft dependency).
• Integration: New metrics are included in console logs and results files.
• Backward compatibility: Default script behavior remains unchanged.

Closes #2506

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[rwightman]

@ha405 looks good, I haven't tested yet, but will when I'm back from vacation in a week and a bit. Did you look at train script? Could be added there but needs extra work to cover distributed case...

[ha405]

hi, I looked into train.py and opened a new request #2574