Infinite-dimensional gradient-based descent for alpha-divergence minimisation