Fix dualtol in HEkkDual::computeExactDualObjectiveValue()

[ambros-gleixner]

In SCIP we sometimes use tighter dual feasibility tolerances; with 1e-10, I discovered this inconsistent use of dual tolerance in computeExactDualObjectiveValue() and the subsequent correctDualInfeasibilities().

Without this fix, the assert

      // no shifts should have occurred
      assert(!ekk_instance_.info_.costs_shifted);

is triggered on instance milo-v12-6-r2-40-1 from MIPLIB 2017.

[jajhall]

Strange: I'll look at this. If I remember correctly, computeExactDualObjectiveValue() is only used to determine the dual objective value with unperturbed costs to determine whether to terminate via the dual bound, so I'm puzzled how the assert is then avoided.

[ambros-gleixner]

The reason is that using the correct (=tighter) tolerance makes computeExactDualObjectiveValue() return minus infinity, so correctDualInfineasibilities() is not called.

[jajhall]

OK I guessed something significant had triggered your interest in this. I need to do some sanity performance checks on latest - since I don't think I've done anything to compromise it - and then I'll merge this.

[ambros-gleixner]

Thanks!