GLM for ill-formed matrices

[olivierlabayle]

For some reason, the GLM package seems to return a NaN value from the first iteration in some occasions. Follow up: https://github.com/JuliaStats/GLM.jl/issues/417. For now, let's just skip those cases but this needs to be investigated further. I suspect many of those issues will be fixed by the next GLM version: either without doing anything or by changing the matrix decomposition to QR (either by default or upon failure)