This fixes an ICE triggered by resolve.c's gfc_expr_to_initialize reseting a
range array ref into a full array ref, updating the rank, but leaving the
shape as is, which eventually leads to an out of bound error.
The right fix would probably be to avoid this kind of tricks. But I don't know
what a patch impleting that would look like.
This patch instead keeps the trick as is. It just frees the shape and re-
resolves the expression, so that rank and shape are updated. It also does a
bit of refactoring about shape freeing.
I think it should be on the safe side, and I'm testing it on x86_64-unknown-
freebsd8.2. OK for trunk if it passes? What about the branches? It is not a
regression, but it looks like a genuine bug.