Data-efficient Kernel Methods for Learning Hamiltonian Systems