[inayatku]

Fixes #590.

The differential variation of Runarsson and Yao updates the k-th ranked survivor with gamma * (x_1 - x_{k+1}), where x_1 is the best survivor. The code had two indices in storage order instead of rank order:

• memcpy(x0, xs, ...) copied storage slot 0, not the best survivor xs + irank[0]*n;
• xs[(k+1)*n+j] is storage slot k+1, not the (k+1)-th ranked survivor xs[irank[k+1]*n+j].

This pull request changes both (variant RB in my comment on #590). The benchmark from that comment, 300 seeds per problem, successful runs and median evaluations:

problem	master	this PR
tutorial	291, 7481	291, 8380
g01	296, 148858	300, 152308
g04	300, 21156	300, 22608
g06	300, 29506	300, 20842
g07	24, 116010	274, 112489
g08	300, 5911	300, 6650
g09	298, 36896	300, 37304
g12	298, 7934	300, 7920
g24	300, 8599	300, 9523
Rosenbrock	300, 5714	300, 5115
Shekel 10	286, 12018	299, 13578

The change makes ISRES more reliable on the harder problems (g07, Shekel 10, g01) and faster on g06 and Rosenbrock. It needs 7% to 13% more evaluations on the tutorial problem, g04, g08 and g24. The benchmark program and the full output are on https://github.com/inayatku/nlopt/tree/isres-590-benchmark/test/isres_benchmark .

If you prefer the smaller change, I can reduce this to the x0 line only (variant B in the comment): it solves g07 in 128 of 300 runs and is at most 4.3% slower than master on these problems.

New test test/t_isres_ranking.cxx: ISRES on g06 with seeds 1 to 10. The runs must all reach f* + 0.1 and need fewer than 250000 evaluations in total. Master: 305863 (fails). This branch: 196685. In a resampling of the 300 benchmark runs, master passes this test with probability 0.005 and this branch fails it with probability 5e-5. The test takes about 1.5 s.

ctest: 89 of 89 pass (MSVC 2022).

[inayatku]

@stevengj, this one is also ready for review. As in the description, I can reduce it to the x0 line only (variant B) if you prefer the smaller change.