Projection-Free CNN Pruning via Frank-Wolfe with Momentum: Sparser Models with Less Pretraining