pgsql: Attempt to fix test interference from foreign_key opfamily test

Attempt to fix test interference from foreign_key opfamily test

The test added by c62b330912e creates a btree opfamily whose members
include built-in integer operators, which get_mergejoin_opfamilies()
finds regardless of schema. So any concurrently running test that
planned an integer join or sort could pick the family up while it
existed. That's the likely cause of the intermittent plan-shape
changes seen in window.sql and equivclass.sql on the buildfarm after
that commit.

Run the test inside a transaction that is rolled back, so the family
is never committed and never visible to other backends.

Per buildfarm members sifaka, prion, turaca, widowbird.

Discussion: https://postgr.es/m/CA+HiwqGitHCmO7nvV4nXXYWDkNYA+zVcJj=JNmL3C1O9B0Gz0A@mail.gmail.com
Backpatch-though: 19

Branch
------
master

Details
-------
https://git.postgresql.org/pg/commitdiff/9e17d25e79d4756be08b4a5521b4b58450217137

Modified Files
--------------
src/test/regress/expected/foreign_key.out | 19 +++++++++++--------
src/test/regress/sql/foreign_key.sql | 19 +++++++++++--------
2 files changed, 22 insertions(+), 16 deletions(-)