unify save/load to policy_spec.json based folders

[relh]

save/load works both on single GPU machine and multi-gpu machine.

---

Unifies checkpoint save/load around policy_spec bundles (directory with policy_spec.json + weights.safetensors) while keeping a minimal .mpt compatibility path via MptPolicy in metta. Adds bundle utilities and a CheckpointPolicy to standardize serialization, resolution, and S3 sync, and fixes submission zip creation to avoid leaking local paths.

Key Changes

• New checkpoint bundle tooling (checkpoint_dir) and CheckpointPolicy for safetensors‑based loading; training/checkpointer now load via policy_spec.
• URI resolution now targets checkpoint directories (and :latest) and handles S3 checkpoint dirs; checkpoint filenames drop the .mpt suffix.
• Evaluator submission zip upload now derives from checkpoint bundle paths and rewrites absolute data paths to portable names.
• CLI/docs/recipes updated to reference checkpoint dirs or policy_spec.json.
• Minimal .mpt compatibility preserved by moving MptPolicy + mpt artifact helpers into metta/rl; .mpt loads only via policy_spec bundles that reference MptPolicy.

Breaking Changes / Migration

• New checkpoints are saved as policy_spec bundles (no .mpt filenames).

[datadog-official]

✅ Tests

🎉 All green!

❄️ No new flaky tests detected
🧪 All tests passed

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 20de671 | Docs | Was this helpful? Give us feedback!}

[nishu-builder]

similarly, can we put all the custom saving/loading into CheckpointPolicy. load_policy_data and CheckpointPolicy.save_policy_data? It would dump (and load) more than just the weights file

[relh]

closing for #4502