Is One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL Training — ThinkLLM