Do not fire useless threads

[jcoupey]

The CVRP::solve implementation fires a number of threads based on the nb_threads parameter (same for VRPTW::solve):

vroom/src/problems/cvrp/cvrp.cpp

Lines 237 to 239 in cd40a64

	for (const auto& param_ranks : thread_ranks) {
	solving_threads.emplace_back(run_solve, param_ranks);
	}

Here each param_ranks refers to a vector of parameters to apply, so each thread is responsible for as many searches that there are parameters in param_ranks.

Possible enhancement

What happens if more threads are provided than searches? For example using -x 0 -t 8, there are 4 searches, so 4 threads out of 8 will run on a empty vector of parameters. One may argue that since the run_solve lambda loops over parameters, nothing is done and it's fine. Fair enough and this is what has been happening ever since the first implementation. Now that would be better still to only fire threads on non-empty parameter sets.

Now a bug

It gets worse with the time-out introduced in #595 since run_solve now divides the available time by the number of parameters:

vroom/src/problems/cvrp/cvrp.cpp

Lines 194 to 200 in cd40a64

	auto run_solve = [&](const std::vector<std::size_t>& param_ranks) {
	try {
	// Decide time allocated for each search.
	Timeout search_time;
	if (timeout.has_value()) {
	search_time = timeout.value() / param_ranks.size();
	}

$ vroom -i input.json -x 0 -t 8
# output OK
$ vroom -i input.json -x 0 -t 8 -l 0
Floating point exception (core dumped)