Change how we pass and resolve supervisor policy uris for behavior cloning

[nishu-builder]

1. Centralized URI-to-PolicySpec resolution. Now metta://xxx works with non-mpt policies

A) The MettaSchemeResolver now returns either:

• The S3 path containing the policy spec (preferred), or
• The raw .mpt checkpoint path for legacy policies

B) policy_spec_from_uri() now handles all URI types uniformly:
- .mpt files → creates MptPolicy spec
- S3 paths → downloads and extracts policy spec archive
- Local dirs → loads policy_spec.json directly

• resolve_uri() renamed to get_path_to_policy_spec_or_mpt() to clarify its purpose

Removed other contextmanager-based load-policy-from-s3 logic, and removed multiple points with custom .mpt handling; now it should all be in policy_spec_from_uri. We now just download to a local cache folder with optional cleanup in atexit handlers

2. Fix passing of behavior-cloning policy URI

Before, we passed around an EnvSupervisorConfig containing a URI string that was being misinterpreted as a path. Now, TrainTool resolves the URI into a PolicySpec upfront and passes the resolved spec directly to VectorizedTrainingEnvironment.

3. Cogames CLI policy parsing cleanup
   The CLI policy parser now uses more shared code for deciding if an input is a URI and parsing it, and allows passing in additional args via key=value pairs.

So now you can do e.g.:

uv run cogames evaluate -m machina_1.machinatrainerbig -p class=random -p metta://policy/dinky,proportion=2 -p s3://softmax-public/policies/daveey.1x4.cvc.dr.maps.1.bct.0/daveey.1x4.cvc.dr.maps.1.bct.0:v1990.mpt,proportion=4

[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: 9137b53 | Docs | Was this helpful? Give us feedback!}