More fun with container types

While poking at the arrays-of-domains TODO item, I found yet another area
in which the existing code fails to handle existing supported cases.
Specifically, although DefineRange will happily take domains or composites
as a range subtype, and we already noticed that domains-over-array-of-
composite are supported, find_composite_type_dependencies is innocent of
the idea that the target type might be embedded in anything but an array
or composite type. The test cases in the attached patch show cases where
HEAD either fails to notice stored data violating a proposed new domain
constraint, or hits assertion failures or core dumps.

The patch addresses this by generalizing the existing special case for
array-over-composite so that any type that's directly dependent on the
target type (which could be its array type, or a domain or range over
it) is treated as a container type and recursively scanned for.

I believe this coding will work without any further changes for the
case of arrays over domains, but I've not fully tested that yet.

I intend to apply and back-patch this shortly, unless anyone has a
better idea about how to handle such cases.

regards, tom lane