prevent random distributions from returning NaN/Inf

[Arzaghi]

Fixes #1174

[MattStephanson]

• For fisher_f_distribution, should we also reject here if _Px == _Ty{1}? That would generate infinity rather than NaN.
  

  STL/stl/inc/random

  Lines 4016 to 4019 in ba6f895

  	_Beta_distribution<_Ty> _Dist(_Vx1, _Vx2);
  	_Px = _Dist(_Eng);
  	return (_Vx2 / _Vx1) * (_Px / (_Ty{1} - _Px));

• The issue mentions student_t_distribution, but there's no change to it. It's not obvious to me how it would cause NaN, but _Rs == 0 would give infinity.
  

  STL/stl/inc/random

  Lines 4145 to 4156 in ba6f895

  	for (;;) { // get a point inside unit circle
  	_Vx1 = _Dist(_Eng);
  	_Vx2 = _Dist(_Eng);
  	_Rs = _Vx1 * _Vx1 + _Vx2 * _Vx2;
  	if (_Rs < _Ty{1}) {
  	break;
  	}
  	}
  	_Rx0 = _CSTD sqrt(_Rs);
  	return _Vx1 * _CSTD sqrt(_Par0._Nx * (_CSTD pow(_Rx0, -_Ty{4} / _Par0._Nx) - _Ty{1}) / _Rs);

[Arzaghi]

Here is my test program with a slight improvement to the @statementreply test code on the issue description.

#include <cassert>
#include <cmath>
#include <random>
#include <iostream>

#include "\tests\std\tests\GH_001017_discrete_distribution_out_of_range\bad_random_engine.hpp"

using namespace std;

template <class Distribution>
void test(Distribution&& distribution) {	
	for (bad_random_generator rng; !rng.has_cycled_through();) {
		const auto x = distribution(rng);
		assert(!isnan(x) && !isinf(x));
	}
}

int main() {
	test(normal_distribution<>{});
	std::cout << "normal_distribution\t\tok" << std::endl;

	test(lognormal_distribution<>{});	
	std::cout << "lognormal_distribution\t\tok" << std::endl;

	test(fisher_f_distribution<>{});	
	std::cout << "fisher_f_distribution\t\tok" << std::endl;

	test(student_t_distribution<>{});	
	std::cout << "student_t_distribution\t\tok" << std::endl;

	std::cout << "Done!";
	std::cin.get();
	return 0;
}

[MattStephanson]

Nice work, this looks mostly good to me. I just have a few minor points.

[cbezault]

If there's something that I'm missing with regards to those static_casts then this LGTM. Otherwise could you just choose a consistent style for getting the type of variables you want?

[StephanTLavavej]

This looks good to me, thanks. I have three superficial comments about comments (heh) and asserts, and one slightly more significant question about using a double versus _Ty constant. If the answer there is that comparing against 1.0e-4 is intentional and correct, I'm happy.

[StephanTLavavej]

Thanks again for fixing this silent bad codegen in <random> - the very worst category of bug! 🪲 ❌ => 😸 🎉