COGNATE: Acceleration of Sparse Tensor Programs on Emerging Hardware using Transfer Learning