Any2Point: Empowering Any-modality Large Models for Efficient 3D Understanding