performance regression in 9.2 when loading lots of small tables

There was a regression introduced in 9.2 that effects the creation and
loading of lots of small tables in a single transaction.

It affects the loading of a pg_dump file which has a large number of
small tables (10,000 schemas, one table per schema, 10 rows per
table). I did not test other schema configurations, so these
specifics might not be needed to invoke the problem.

It causes the loading of a dump with "psql -1 -f " to run at half the
previous speed. Speed of loading without the -1 is not changed.

The regression was introduced in 39a68e5c6ca7b41b, "Fix toast table
creation". Perhaps the slowdown is an inevitable result of fixing the
bug.

The regression was removed from 9_1_STABLE at commit dff178f8017e4412,
"More cleanup after failed reduced-lock-levels-for-DDL feature". It
is still present in 9_2_STABLE.

I don't really understand what is going on in these patches, but it
seems that either 9_1_STABLE now has a bug that was fixed and then
unfixed, or that 9_2_STABLE is slower than it needs to be.

The dump file I used can be obtained like this:

perl -le 'print "set client_min_messages=warning;"; print "create
schema foo$_; create table foo$_.foo$_ (k integer, v integer); insert
into foo$_.foo$_ select * from generate_series(1,10); " foreach
$ARGV[0]..$ARGV[0]+$ARGV[1]-1' 0 10000 | psql > /dev/null ; pg_dump >
10000.dump

Cheers,

Jeff