Stabilize ordered_append by widening device count

The CTE query in ordered_append-* sat at the cost crossover between
Index Only Scan via device_id_time_idx and Index Scan via time_idx on
metrics_space chunks. With one device per chunk after hash partitioning,
IOS had no selectivity advantage at the chunk level, so any small
relallvisible drop tipped the plan choice. Bumping the metrics_space
pair to 20 devices gives ~7 devices per chunk, restoring IOS's
structural advantage and making the choice robust to VM noise.

Author: antekresic