Gentle Manipulation Policy Learning via Demonstrations from VLM Planned Atomic Skills