Averaging of PSFs with different wavelength ranges

[marcelo-alvarez]

Attempt to fix bugs preventing PSF and trace coefficients from being correctly averaged in desispec.specex.mean_psf whenever the wavelength range differs between the input PSF fits:

1. legval and legfit were not being imported from numpy.polynomial.legendre
2. traces were being averaged redundantly for every entry in the table of PSF coefficients
3. traces were being re-fit using the wrong number of legendre polynomial coefficients

These changes may require further validation; there is no before and after comparison possible for the use case that has changed since it was crashing for these cases (i.e. whenever the wavelength range differs between the input PSF fits).

[coveralls]

Coverage decreased (-0.002%) to 26.667% when pulling 5fabc5d on meanpsf into e767d60 on master.

[julienguy]

Thanks for finding those bugs.
Tested on June data.