Dear Fortranners,

the fix for initialization of DT arrays caused an apparent regression for
cases where inconsistent ranks were used in such an initialization.
This caused either an ICE in subsequent uses of these arrays, or showed
up in valgrind as invalid reads, all of which seemed to be related to this
rank mismatch.

The cleanest solution seems to be to strictly reject rank mismatch earlier
than we used to, which helps error recovery.  I had to adjust one testcase
accordingly.

The place I inserted the check does not distinguish between explicit shape
and implied shape.  The Intel compiler does give a slightly different
error message for the implied shape case.  If anyone feels strongly about
this, I'm open to suggestions for better choices of handling this.

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

Thanks,
Harald