Work around for ev11 crash

[phyy-nx]

Work around for when ev11 gets a negative global_sum_of_delta_squared_in_bin for a given bin.

The test case I have shows that the refinement takes a bad step giving a high target function value, even if I throw out a bad bin. It then recovers the minimization procedure:

5669.41
5669.41
2638.52
2571.27
2291.39
1115.96
2253525.53  <-- this is the bad step with the single bin that needs to be thrown out to avoid a math domain error
1115.78
2289.79
943.97
936.21
927.23
918.19
868.21
803.22
799.45