get_eclass_for_sort_expr() computes expr_relids and nullable_relids
early on, even though they won't be needed unless we make a new
EquivalenceClass, which we often don't.  Aside from the probably-minor
inefficiency, there's a memory management problem: these bitmapsets will
be built in the caller's context, leading to dangling pointers if that
is shorter-lived than root->planner_cxt.  This would be a live bug if
get_eclass_for_sort_expr() could be called with create_it = true during
GEQO join planning.  So far as I can find, the core code never does
that, but it's hard to be sure that no extensions do, especially since
the comments make it clear that that's supposed to be a supported case.
Fix by not computing these values until we've switched into planner_cxt
to build the new EquivalenceClass.

generate_join_implied_equalities() uses inner_rel->relids to look up
relevant eclasses, but it ought to be using nominal_inner_relids.
This is presently harmless because a child RelOptInfo will always have
exactly the same eclass_indexes as its topmost parent; but that might
not be true forever, and anyway it makes the code confusing.

The first of these is old (introduced by me in f3b3b8d5), so back-patch
to all supported branches.  The second only dates to v13, but we might
as well back-patch it to keep the code looking similar across branches.

Discussion: https://postgr.es/m/1508010.1601832581@sss.pgh.pa.us