Re: Oversight in reparameterize_path_by_child leading to executor crash

I spent some time reviewing the v4 patch.  I noted that
path_is_reparameterizable_by_child still wasn't modeling the pass/fail
behavior of reparameterize_path_by_child very well, because that
function checks this at every recursion level:

    /*
     * If the path is not parameterized by the parent of the given relation,
     * it doesn't need reparameterization.
     */
    if (!path->param_info ||
        !bms_overlap(PATH_REQ_OUTER(path), child_rel->top_parent_relids))
        return path;

So we might have a sub-path (eg a join input rel) that is not one of
the supported kinds, and yet we can succeed because parameterization
appears only in other sub-paths.  The patch as written will not crash
in such a case, but it might refuse to use a path that we previously
would have allowed.  So I think we need to put the same test into
path_is_reparameterizable_by_child, which requires adding child_rel
to its param list but is otherwise simple enough.

So that led me to the attached v5, which seemed committable to me so I
set about back-patching it ... and it fell over immediately in v15, as
shown in the attached regression diffs from v15.  It looks to me like
we are now failing to recognize that reparameterized quals are
redundant with not-reparameterized ones, so this patch is somehow
dependent on restructuring that happened during the v16 cycle.
I don't have time to dig deeper than that, and I'm not sure that that
is an area we'd want to mess with in a back-patched bug fix.