Bug fix for mode_compute=6.

@stepan-a I would have preferred to discuss this change before pushing. It implies a big change in terms of backward compatibility. Instead of computing the covariance from scratch based on the existing sample draws, we now update gradually based on an initialization.

Ah... My understanding is that it fixes the bug you spotted. As far as I am concerned, I always had in mind that the initial variance should have a (small) effect and that repeated calls to gmhmaxlik_core should keep track of the current values for the mean and covariance matrix. It's true that the patch changes the results. I'll revert the cherry-pick in 5.x.