Assertion fail in tanh_sinh when using cpp_dec_float_100

[mjziebarth]

Hi all,

this might be closely related to #155. I'm not sure about your issue policy, so feel free
to close this one and reopen the old one.

I have encountered a similar issue, triggering the Assertion `(left_min_complement * diff + a) > a' failed error
when computing an integral using a higher precision boost multiprecision number type. After some investigation,
I realized that the error occurs before the first call of the integrand, and I could boil down the problem to the
integration bounds. When choosing an integration bound slightly lower than 1, the error occurs.

I found out that the error can be relaxed by either choosing 1 as an integration bound and when switching to
a native floating point type. Here's a minimal example that, on my machine, fails when compiled without further
defines and succeeds with NO_BUG or ALSO_NO_BUG defined.

#include <boost/math/quadrature/tanh_sinh.hpp>
#include <boost/multiprecision/cpp_dec_float.hpp>

using boost::math::quadrature::tanh_sinh;

#include <iostream>

int main(){
#ifdef ALSO_NO_BUG
	typedef long double real;
#else
	typedef boost::multiprecision::cpp_dec_float_100 real;
#endif

	auto fun = [](real x) -> real {
		return 1.0;
	};

	tanh_sinh<real> integrator;
	const real a = 0.0;
#ifdef NO_BUG
	const real b = 1.0;
#else
	const real b = 0.9999995515592481132478776023609116290187750667053638330158486516399489191171270344610533516275817076;
#endif
	real y = integrator.integrate(fun, a, b);
	std::cout << std::setprecision(10) << y << "\n";

	return 0;
}

In case it's relevant, I am running an Arch linux system with boost 1.80.0-3 installed.