feat(rewrites): improve how constant sort values are dropped

[NickCrews]

Before, we only dropped constant sort values if

• it was an EXACT Literal. if it was derived, eg ibis.literal(5) + 1, then we wouldn't.
• if the merge_select_select decided to return the original select instead of merging them (eg because of a window function, unnest, etc), then we wouldn't see this optimization.

This is a followup to a TODO that I found at 7a6f0a1#diff-2719b402427dca65945905bfc259790600013287258e555367581729c40d837eR323-R328

[cpcloud]

Why is there now a subquery? Is that correct?

[NickCrews]

Did you see how the actual expression changed here? Before the expression was just t.order_by(ibis.random()). Now it is the more complicated t.order_by(ibis.random()).select("f", "c").order_by(ibis.random()).

This new version is equivalent to t.select("f", "c").order_by(ibis.random()) which is SELECT "t0"."f", "t0"."c" FROM "star1" AS "t0" ORDER BY RANDOM() ASC. For this simple case we could drop the internal order by, and then merge the two selects. But my initial gut reaction is that there are cases where the internal select result depends on the internal order by, and so in general we can't just drop the inner select? So in general I would say this subquery is unnecessary, but correct.