Incentivizing Dual Process Thinking for Efficient Large Language Model Reasoning