I've tried benchmarking the changes here and the difference is not noticeable because:

• instances for which there are several similar heuristic solutions are mostly the easy/small ones so convergence is already super fast on those;
• medium to big instances seldom have similar heuristic solutions, even with -x 5;
• removing a couple local searches does not usually impact overall computing time in a context where multiple threads are used.

So I don't think it's actually worth adding this as is since:

• it adds quite some boilerplate to the solving code in cvrp.cpp and vrptw.cpp, splitting the multi-threading part;
• the current solution comparison based on global indicators can also "shadow" some solutions that look identical but for which actual routes are different; in some cases this actually cuts unwanted searches, sometimes leading to a slightly poorer overall solution.