Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders