Symplectic Spectrum Gaussian Processes: Learning Hamiltonians from Noisy and Sparse Data