GLM: Fix all_stats() segfault

[Lestropie]

As reported in #2289.

Fixes segmentation fault in case where:

1. Either:
   1. Input data contain non-finite values (regardless of whether or not those data are or are not within the statistical inference processing mask);
   2. Element-wise design matrix columns are added via -column option;
2. At least one F-test is performed.

Fairly obvious blunder even without explaining the surrounding code.

[jdtournier]

I'm not sure I fully appreciate what each row of your local_abs_effect_size and local_std_effect_size is supposed to contain, but I'll take your word for it that it's appropriate to set the first row to NaN rather than the second... 😁

[Lestropie]

I'm not sure I fully appreciate what each row of your local_abs_effect_size and local_std_effect_size is supposed to contain ...

In cases where the GLM can't be fit to the exhaustive set of data in a single matrix operation, the relevant code loops over elements, each thread processes one element at a time. So any GLM derivative that would normally have one row per element becomes a row vector in that context. But they're treated as matrices with one row rather than explicit vectors so that the same function can be used as that used when all data are solved at once.