Towards Understanding Adversarial Transferability From Surrogate Training