Fix ICLS performance regression

[daljit46]

Fixes #3260.
The following change was introduced in #3182:

-    std::fill(active.begin(), active.end(), false);
-    if (num_eq > 0)
-      std::fill(active.begin() + num_ineq, active.end(), true);
+    active.setZero();
+    active.tail(num_ineq).setOnes();

We were mistakenly setting the last num_ineq entries active, while we should be doing the opposite: set the last num_eq entries active. This caused a large set of inequality constraints to be active in the solver, leading to performance issues. This PR restores the behaviour prior to #3182.

[github-actions]

clang-tidy review says "All clean, LGTM! 👍"

[github-actions]

clang-tidy review says "All clean, LGTM! 👍"

[bjeurissen]

Well spotted! Was it only a matter of speed or did it also change the outcome? I would suspect the former as the tests still passed.

[jdtournier]

As far as I can tell, that basically meant it was initialised with all constraints active, and since constraints are added/removed one at a time, this means it takes a lot longer to converge to the solution (which will typically have only a handful of constraints active). That explains the longer runtime, and why all the tests passed. Nice detective work!

[jdtournier]

Fix looks good to me.