Erase or Hide? Suppressing Spurious Unlearning Neurons for Robust Unlearning