Taming the Sigmoid Bottleneck: Provably Argmaxable Sparse Multi-Label Classification