Fixed stability issues in graph pyramid

[EiffL]

First of all, thanks so much for making these tools public. I encountered just a few issues when using pygsp.operators.reduction.graph_multiresolution in practice:

• If a Graph object is fed to graph_sparsify instead of a weight matrix, there is no thresholding of small values, which lead in practice to nans in the resulting matrix
• also in grap_sparsify a parameter q is computed, should be an int except for me it wasn't
• in kron_reduction I encountered some numerical stability issues and non-zero diagonals after reduction, not 100% sure it is correct to do so but I just added a line to remove the diagonal of the resulting matrix

[rodrigo-pena]

Thanks a lot for your contribution. I'll merge it to the master branch. Feel free to contribute whenever you like.