A dynamic generative model can extract interpretable oscillatory components from multichannel neurophysiological recordings