How to Scale Inverse RL to Large State Spaces? A Provably Efficient Approach