Stop when encountering a NaN

[pjotrp]

When we see:

 ## REMLE estimate for Ve in the null model:
 2.14138e-05
 -3.21985e-06    1.68684e-06
 -1.21203e-05    -1.80521e-06    1.82497e-05
 ## se(Ve):
 1.72819e-05
 5.62131e-06     2.21151e-06
 1.07007e-05     -nan    1.17464e-05

there is a nan which suggests the results are unstable. We are
dealing with round-off errors most likely (division of values close to zero).

Shouldn't gemma stop processing when it detects such a situation.
In fact I want to suggest GEMMA stops at any point a nan gets
introduced anywhere because it implies we are operating at machine
boundaries and results can not be trusted.

Xiang wrote:

Yes, I agree. I have seen this nan error a few times before, in either the Ve or Vg estimates. Sometimes better phenotype normalization can help preventing the error, but sometimes doesn't -- and in those cases I couldn't figure out why. But I agree it makes sense to stop processing when it detects such an error.

[pjotrp]

We now have default hardware detection of FP out of boundary errors on Intel x64. Releasing soon.