Fix: Block deferred-computation exploit and harden validator (#64)

* Fix: Block deferred-computation exploit and harden validator

- AST scanner: reject forbidden dunder method definitions (e.g. def __getattribute__)
  that bypassed the existing ast.Attribute check
- _validate_return_type: expand MRO proxy detection to cover __getattribute__,
  __get__, __set__, __set_name__ in addition to __getattr__
- Add timing guard on final_state access to detect training compute
  deferred into state gathering (rejects if >50% of training wall_time)
- Harden final_state dict-value proxy detection with same expanded hook set
- Basilica polling: reuse single httpx client, increase read timeout to 60s,
  dynamic max_consecutive_errors based on remaining time budget
- Set cudnn.deterministic=True, cudnn.benchmark=False for reproducibility
- Remove wall_time rank divergence check (use min across ranks instead)
- Bump docker memory in simulate_validator.py to 400g

Made-with: Cursor

* Fix arbos container cleanup killing unrelated containers

The _post_run_cleanup in LocalDockerTester was killing ALL templar-eval
containers by filtering on ancestor image. This caused manual test runs
and other instances to be SIGKILL'd (exit 137) mid-execution. Now each
arbos session labels its containers and only cleans up its own.

Also remove unnecessary --memory and --shm-size from simulate_validator
docker examples (--ipc=host already handles shared memory, and the memory
limit was not the cause of the crashes).

Made-with: Cursor

* Align arbos BasilicaTester with production runner

Switch from low-level create_deployment_async to deploy_async,
add log streaming, result validation, proper exception handling
(DeploymentTimeout/DeploymentFailed), and update stale defaults.

Made-with: Cursor

Author: shivam-MBZUAI