Make lambda a vector instead of a scalar

[fjandrad]

This can further enhance the regularization, so that we can include a degree of confidence in the reference value.
For example if the z estimation is not so bad we could increase the lambda in the z axis instead if it is bad we can decrease so that we trust more the new fitting than the previous fitting.

This could help since usually the big errors occur in the y axis while when in 2 feet the estimation from the workbench in the z axis seems not so bad.

[traversaro]

This make sense.

[fjandrad]

When estimating using just the calibration matrix from the workbench as previous is not so relevant, since on the secondary matrix test selects by itself the best lambda.

It is worth exploring in the case where there are other types of previous information such as offset information.