Lottery Tickets in Evolutionary Optimization: On Sparse Backpropagation-Free Trainability