Provable in-context learning of PDEs