gfortran.dg/gamma_(1|2).f90 fail on Darwin

[Dominique Dhumieres]

> I propose the following patch

Seems reasonable to me; but for the new test if "fortran_large_real"
is "true" for systems with real(16): it will give the same failure
since the exponent range is only 291. I think
selected_real_kind (precision (0.0_8) + 1)
is better.

> (cannot test nor commit, sorry)

I can test, but not commit, although I am pretty sure that the patch is OK
excepted for gamma_4.f90.