ALTER COLUMN TYPE vs. domain constraints

I found out that altering a column's type does not play nicely with
domain constraints: tablecmds.c expects that only table constraints
could depend on a column. Now, it's easy to hit that with domains
over composite, so I propose to fix it in HEAD with the attached
patch. However, if you really work at it, you can make it fail
in the back branches too:

regression=# create type comptype as (r float8, i float8);
CREATE TYPE
regression=# create domain silly as float8 check ((row(value,0)::comptype).r > 0);
CREATE DOMAIN
regression=# alter type comptype alter attribute r type varchar;
ERROR: cache lookup failed for relation 0

Before commit 6784d7a1d, the ALTER actually went through, leaving a
mess. Fortunately it doesn't actually crash afterwards, but you
get things like

regression=# select 0::silly;
ERROR: ROW() column has type double precision instead of type character varying

We could consider back-patching the attached to cover this, but
I'm not entirely sure it's worth the trouble, because I haven't
thought of any non-silly use-cases in the absence of domains
over composite. Comments?

regards, tom lane