Replace custom AsyncEvaluator with torchrl Evaluator

[vmoens]

Summary

• Replaced the hand-rolled AsyncEvaluator class and run_eval/collect_eval_result/log_eval_result helpers with the unified Evaluator from torchrl.collectors
• The torchrl Evaluator handles weight transfer, video dumping, and metric logging internally for both sync (evaluate()) and async (trigger_eval()/poll()) modes
• Removed queue, threading, ExplorationType, set_exploration_type imports that were only used by the removed code

Test plan

• Run humanoid PPO training without compilation to verify basic functionality
• Run humanoid PPO training with compilation on both collector and evaluator envs
• Verify eval metrics and videos are logged to WandB correctly

🤖 Generated with Claude Code