Thinking-Free Policy Initialization Makes Distilled Reasoning Models More Effective and Efficient Reasoners