keep the sign of cos(y) in sinh near the overflow boundary - #361

#352 moved the sign of x into the phase on the large |x| path of sinh , but it does it with RealType(signOf: z.x, magnitudeOf: RealType.cos(z.y)) , which also throws away the sign of cos(y) . so whenever cos(y) is negative the real part has the wrong sign , and sin , which goes through sinh , gets the wrong imaginary part for a large imaginary input :

Complex<Double>.sinh(Complex(40, 3 * .pi/4))
// main  (8.32212591859309e+16, 8.322125918593091e+16)
// fixed (-8.32212591859309e+16, 8.322125918593091e+16)

Complex<Double>.sin(Complex(.pi, 40))
// main  (14.413190679920108, 1.1769263341851e+17)
// fixed (14.413190679920108, -1.1769263341851e+17)

the fixed values match mpmath . the fix multiplies cos(y) by the sign of x , the same way cosh already does with sin(y) .

the tests from #352 only use y = ±π/4 , where cos(y) is positive . this adds the same near overflow checks at y = ±3π/4 for sinh , and two for cosh so both functions cover that quadrant . the four new sinh real part checks fail on main for Float and Double , and the whole suite passes with the change .

i found it by comparing the complex elementary functions against mpmath on random inputs over a wide range of magnitudes . after this change everything else was within 5 ulp normwise , apart from a few tiny atan / atanh inputs where mpmath itself ran out of precision .