normalize residuals in contrib_gencoeffs to help solver with poor initial guesses

[mikofski]

relevant to #88, currently the gen_coeffs solver is very sensitive to initial guess, and so any non sunpower module is hard to solve for. when using PVLife we never had this problem, but in PVLife, all of the parameters were scaled to values between 1 and zero, this helps the solver.

EG: use

x[0] = log(isat1) / log(isat1_char)
x[1] = log(isat2) / log(isat2_char)
x[2] = (rs / rs_char)**2
x[3] = (rsh / rsh_char)**2

where _char means a "characteristic" value for the class of pv device being fit.

[markcampanelli]

@mikofski Again I make no claims about the effectiveness of my approach as compared to others, but I normalize single I-V curves against the "empirical" Isc and Voc values from the data, and after some algebra this will give a rescaling of all the parameters too. For sets of I-V curves (e.g., an IEC-81653 "matrix" of curves), I scale all the data to the empirical STC values Isc0 and Voc0, which assumes that there is a curve in the set that is sufficiently close to STC to work with.