See if the WavelengthCalibration1D.refine_fit penalty function can be made more robust

[hpparvi]

@kbwestfall noticed that the penalty function used to refine a wavelength calibration solution in WavelengthCalibration1d may lead to situations where the fit is improved by finding solutions that exclude lines rather than improving the polynomial. We need to take a look at how to make the penalty function more robust.

#265 (comment)