only solve problems if have enough precision

[ryanelandt]

This is chunk 3 for #1000

In the ibeta_inverse.hpp file, there exists root-finding problems of the form:

log(x)  + a * log(1 - x) + t

This function has an extrema at 1 / (1 + a).

Newton solver finds dubious roots

For extreme values of a, the extrema evaluates to 0 or 1 due to floating point limitations. For these cases, the objectives function evaluation always has the same sign, and is never zero. The Newton solver finds a root for the reasons described in #1005.

This PR

• Uses if (cross == 0 || cross == 1) { return cross; } to return early for values that lack the numerical precision to construct a meaningful objective function
• Adds a BOOST_MATH_ASSERT to prevent the construction of root finding problems that are problematic.

[mborland]

I'll look into the CI system to see what's going on with the TSAN failures across all of your PRs, but it is clearly unrelated.