Remove approximate indexes for ORDER BY queries without a LIMIT - #1031

Hi @oliness, thanks for the PR.

This seems to work well for some queries I've tested (better than limit_tuples in #424), but I'm worried it's more risky / brittle / could break other queries. For instance, tuple_fraction is set to zero in a number of situations.

@jkatz @hlinnaka any thoughts?

I've added TAP tests for the tuple_fraction = 0 case: test/t/049_hnsw_no_limit.pl and test/t/050_ivfflat_no_limit.pl, one per index type. They run with enable_sort = off, so the index is chosen whenever it's allowed.

For queries planned with tuple_fraction = 0, they check that the index isn't used and that the results match a scan with enable_indexscan = off:

• no LIMIT
• a subquery without its own LIMIT, under an outer aggregate, GROUP BY, ROLLUP, HAVING, DISTINCT, ORDER BY on another column, or join
• the same kind of subquery when the outer query only reads the first rows: an outer ORDER BY on the same distance with a LIMIT, or a join ordered by distance with a LIMIT

They also check that these queries still use the index:

• a LIMIT
• an outer LIMIT with nothing else in the outer query
• a subquery with its own LIMIT, under an aggregate, ORDER BY, or join
• no LIMIT with enable_seqscan = off, at the top level and in a subquery

On master, 17 of the 27 checks in each file fail, including every "no index scan" check. The checks for queries that should still use the index pass on master too.