Use threads more effectively in differential evolution

[NAThompson]

No description provided.

[NAThompson]

Using this code:

void test_sphere() {
  std::cout << "Testing differential evolution on the sphere function . . .\n";
  using ArgType = std::vector<float>;
  auto de_params = differential_evolution_parameters<ArgType>();
  de_params.lower_bounds.resize(12, -1);
  de_params.upper_bounds.resize(12, 1);
  de_params.NP *= 150;
  de_params.max_generations *= 10;
  de_params.crossover_probability = 0.9;
  double target_value = 1e-8;
  std::mt19937_64 gen(56789);
  auto local_minima = differential_evolution(sphere, de_params, gen, target_value);
  CHECK_LE(sphere(local_minima), target_value);
}

I obtain:

./a.out  11.57s user 0.30s system 113% cpu 10.430 total

on develop, and

./a.out  14.23s user 0.31s system 564% cpu 2.574 total

on top of this PR. So it appears this is an unambiguous win-though in fairness my laptop has 10 threads so there is much performance left to be extracted.

The PR is clean under -fsanitize=thread as well as -fsanitize=address -fsanitize=undefined.