Moccasin: Efficient Tensor Rematerialization for Neural Networks