CRSMatrixDescriptor with large rows

The access cost to an (i,j) element of the matrix descriptor depends linearly on the number of elements in the ith row (wort case scenario).

In the case of large rows(>>10 values?), this linear search could be assisted by a bisection method to predict the location of the column. The idea (as discussed with @hpwxf) could be to perform this bisection until the remaining set of column number candidates is smaller than a prescribed value (10?), and then to finalize the search using the current method (array browsing).