Make the sparse qr solver accept a diagonal conditioner

[furgalep]

This may not make sense for the incremental calibration project but there is nothing logically wrong with it. Currently we are implementing a small class for trust region methods and this would allow us to run SPQR with LM, for example.