Efficient Neural Theorem Proving via Fine-grained Proof Structure Analysis