NLA affect training checkpoints

Resumable training state for the NLA affect experiments.

Layout:

• av/latest/step_000073: latest resumable AV state after Exp19.
• ar/latest/step_000209: latest resumable AR state after Exp18.
• logs/: metrics, status, split manifests, and launch logs for the latest runs.

Each rank state contains model, AdamW optimizer, scheduler, gradient-scaler/RNG state, and is intended for exact three-rank resume with the corresponding source code. Older intermediate checkpoints are deliberately not archived. The portable inference checkpoints are in the AV and AR model repositories, not here.