STEP: Learning N:M Structured Sparsity Masks from Scratch with Precondition