Add NeuTraL-AD model, loss, detector, transform and experiments

[denix56]

Motivation

• Integrate the NeuTraL-AD (Neural Transformation Learning for Anomaly Detection) method so it can be trained and evaluated within TimeSeAD.
• Provide a lightweight filter to select windows by label to support NeuTraL-AD’s self-supervised training pipeline.
• Add experiment code and configs to enable grid-search / training runs for SMD and reconstruction benchmarks.
• Document the third-party AGPL license attached to the NeuTraL-AD source file to inform users of licensing constraints.

Description

• Added the NeuTraL-AD implementation in timesead/models/other/neutral_ad.py including NeutralAD model, NeutralADLoss, and NeutralADAnomalyDetector.
• Exposed the new classes in timesead/models/other/__init__.py and added a training experiment timesead_experiments/other/train_neutral_ad.py that integrates with the existing training_ingredient and data_ingredient and uses instantiate_loss to construct the loss.
• Implemented a WindowLabelFilterTransform in timesead/data/transforms/target_transforms.py and exported it via timesead/data/transforms/__init__.py to allow filtering windows by label (e.g., keep normal-only windows during pretraining).
• Added experiment configurations experiment_configs/smd/other/train_neutral_ad_on_smd.yml and experiment_configs/recon/other/train_neutral_ad_recon.yml, and updated README.md with an AGPL licensing notice for the third-party code.

Testing

• No automated tests were executed for this change.
• Basic repository sanity (file additions and imports) were committed successfully to the local repository.
• Users should run the existing experiment grid or unit tests in their CI to validate training and integration on target environments.
• Recommend running timesead_experiments workflows and unit tests that exercise timesead.data.transforms and timesead.models.other on CI to verify runtime behavior.

---

Codex Task