bilby/gw/likelihood.py: make phase of d_inner_h consistent with that of the other likelihoods

d_inner_h for multiband was np.conj(data).dot(template) but for the other likelihoods np.conj(template).dot(data). This MR takes the complex conjugate of the multiband d_inner_h to make them all consistent. Since likelihood depends on its real part (absolute part for phase marginalization), this difference does not bias parameter recovery results during sampling. However, phase resampled from phase marginalization will be mapped into its negative. This fixes #678 (closed).