Sgap: Towards Efficient Sparse Tensor Algebra Compilation for GPU