Subtract constant terms (in Q) so that action Matsubara summation/energy
integration becomes convergent at the saddle point.

Since we are currently only using the Q-derivatives of the action, this
does not change anything.

However, for calculation of the free energy, this subtraction is useful
for cutoff elimination, as it replaces the mean field term

    |Delta|^2 / lambda -> log(T/T_c) |\Delta|^2

so that it is only defined by the weak-coupling critical temperature.

This is free energy plus some offset.

Work around apparent bug in scipy.linalg.signm() in some cases computing
the wrong result.

cppad.general is much faster than cppad.symmetric.

Also computing only lower triangle subset speeds things up slightly.

Selfcons iteration doesn't need full solver reset, only drop the
Jacobian decomposition.