feat(openclaw-rl): add reward model feedback training

[chengzidl]

Summary

• add a persistent feedback store for implicit positives and thumbs-down negatives
• add a learned reward model with background training on collected feedback
• integrate reward-model scoring into openclaw_api_server.py
• expose /v1/feedback for thumbs-down collection
• document and enable reward-model env vars in the launch script

Notes

• thumbs-down now rewrites the stored label for the same turn instead of duplicating it
• kept unrelated local files out of the commit

[yinjjiew]

Why do we need to add another reward model and keep training it here? Isn't using a PRM to judge response quality based on next-state environment/user feedback enough in this setting? I think most users who deploy personal agent don't have so many computations.