Optimize coalesceSingle

[yhahn]

Adds some non-invasive optimizations to coaleseSingle taking advantage of the fact that covers are sorted in relev, score descending order. Since the cutoff for coalesce is 40 features, we can also short circuit intelligently based on this cutoff.

• When not using proximity or bbox, we can short circuit after the 40th unique feature ID is collected
• When using proximity/bbox, we need to iterate further but can know intelligently when we've reached a cutoff point where no further iteration will yield more useful features, because:
  • Covers are sorted by relev first. If the relev of subsequent features in the loop drops below the smallest relevance collected and we have 40 features already, we're done

I've tested this with carmen + IRL scenarios to consider it ready for tagging.

Benchmark numbers

IRL benchmarks are more meaningful but for the synthetic coalesceSingle() scenarios:

master

# coalesceSingle
ok 1 coalesceSingle @ 1.88ms
# coalesceSingle proximity
ok 2 coalesceSingle + proximity @ 2.26ms

early-death

# coalesceSingle
ok 1 coalesceSingle @ 0.24ms
# coalesceSingle proximity
ok 2 coalesceSingle + proximity @ 0.38ms