Consider to a sparse-sparse matmul case

[cshjin]

Overall, if we have two sparse matrices, then the multiplication should be sparse as well, in order to improve the performance.
Example like https://github.com/scipy/scipy/blob/a03f15fe43cb68c132192af22deb73b69ef201bf/scipy/sparse/compressed.py#L496-L530
And the spspmm function in the pytorch-sparse package. https://github.com/rusty1s/pytorch_sparse/blob/master/torch_sparse/spspmm.py

While in TF2.x, no matter the matmul (sparsetensor, sparsetensor) or the sparse.sparse_dense_matmul(sparsetensor, tensor) function return a Tensor instead of SparseTensor.

This is the one of the reasons why the PyG is much faster in the implementation of tensorflow or DGL.

Performance issue need to be resolved.