odd behaviour of multifit and covariance matrix

[abuts]

on some machines at some rng values the behaviour of multifit becomes ill defined as covariance matrix calculated at the last step of fitting become singular.

The remidy for that is to use normalized covariance, (See reference to this ticket in the code, .\Horace\herbert_core\applications\multifit\@mfclass\private\private\multifit_lsqr_ser.m) which is used during fitting process, but this gives much higher estimate for the fitting errors.

The ticket is to estimate correct approach for calculating the covariance and error during the fitting process.