Learning Orthogonal Multi-Index Models: A Fine-Grained Information Exponent Analysis