tp: use HLL cardinality estimates in query planner

Wire up Column::estimated_unique_count into the query planner's
equality filter row count estimation. When an HLL estimate is
available, use N/unique_count instead of the old N/(2*log2(N))
heuristic. This gives SQLite more accurate cost estimates, leading
to better join ordering.

Fix CUJ stdlib tables to use slice.id instead of ts for
row_number() ordering. slice.id is equivalent to ts ordering
(slices are inserted in ts order) but is resilient to ties when
multiple CUJs share the same timestamp.

Fix cpu_cycles_per_thread_slice diff test to add ORDER BY id
for deterministic output with LIMIT.