Spurious underflow in cyl_bessel_k

[jzmaddock]

cyl_bessel_k(100, 746) returns zero when evaluating at double precision (with no extended long double type), but the result should be 9.49147e-66.

Underflow occurs in k0 and k1 prior to iteration. The problem is also present if v is perturbed to be a non-integer in which case CF2_ik underflows.

Reduced test case from #1228.

[NAThompson]

@jzmaddock : On wolfram alpha I get ~3e-323, which is a denormal.

[jzmaddock]

Correction, try:

cyl_bessel_k(-1000, 746)

We underflow to zero, wolfram gives 3.365e-67.