The Counterattack of CNNs in Self-Supervised Learning: Larger Kernel Size might be All You Need