[v0.2] Add lazy import to fix circular import and ray init config support

[listar2000]

Description

This PR merges two issues encountered when testing the examples/math_tools/train_math_with_tool.sh script.

1. Circular import issue between engine and workflows modules
   Previous cycling pattern:
   workflows/__init__.py → workflows/workflow.py → engine/rollout/rollout_engine.py → engine/__init__.py → engine/agent_workflow_engine.py → workflows/workflow.py ← CYCLE!

2. Improves Ray initialization (ray.init) configuration to allow more arguments other than num_cpus.

   Now it's easier to specify arguments like (_temp_dir, address, logging_level) under the ray_init fields.

Changes Made

• Implemented lazy import for AgentWorkflowEngine in engine/__init__.py using __getattr__() pattern (similar to workflows/__init__.py.
• Added None value filtering in agent_trainer.py before passing config to ray.init()

Motivation

The circular import prevented the engine module from being imported successfully, causing the training to fail on initialization. The ray configuration improvements allow users to properly configure Ray's temporary directory and other settings through Hydra configuration.

Testing

• Pass the pytests (a couple of seemingly unrelated errors in tests/agents/test_math_agent.py)
• Tested locally (with examples/math_tools/train_math_with_tool.sh) with multi-GPU training setup
• Verified circular import is resolved
• Confirmed Ray initialization works with custom configurations (e.g. _temp_dir)