ENH: overhaul of singlet fitting methods

[bennyrowland]

The current model for singlet fitting uses a VARPRO approach where the linear components (the peak amplitudes) are fitted in a linear least squares step independently from the non-linear (width, frequency, phase) components. This is more efficient because no iteration is needed for those parameters, and it reduces the size of the non-linear search space, however it also makes constraints between the peak amplitudes impossible to introduce, and also gives challenges in returning fit standard errors because of the split in method. I propose to modify the singlet fitting code to use entirely non-linear methods - this will allow much greater alignment with the standard lmfit template and so give more flexibility to the fitting, allowing different optimisation algorithms to be specified, for example.