Regressions in elliptic_pi_inc

[fredrik-johansson]

With the new algorithm to evaluate R_J, there are some regressions in elliptic_pi_inc (see https://trac.sagemath.org/ticket/29719):

• Slightly lower accuracy
• NaN result when evaluated at inexact phi = pi/2 (this corresponds to an inexact zero as input to R_J)

Either further analysis could be done for the Carlson R_J algorithm to establish larger regions of validity, or perhaps a separate algorithm could be added to deal with the inexact phi = pi/2 case.