Dear all,

it appears that the fix for pr107874 uncovered a latent bug
for the case of arrays of type character and size zero when
passed to the intrinsics MERGE and SPREAD as SOURCE.  In that
case, there is no constructor from which we could obtain
another character length.  A reasonable solution seems to
retain the array's character length.

Since I could not find a simple case where this fails, I've
added an assertion that we actually have a meaningful length.

Regtested on x86_64-pc-linux-gnu.  OK for mainline?

Thanks,
Harald