Avoid redundant work when writing features

[geoneutrino]

small performance improvement

Two pre-existing redundancies in the tile write path, neither of which changes what is written:

• writeMultiLinestring() called scaleLatpLon() twice for every point, once to count the distinct points and once to emit them. Scale each point once into a reused buffer instead.
• The multipoint vector was allocated per point feature. Reuse a thread_local buffer, as scaledMultiPolygon already does.

Verified byte-identical to the parent commit on a Bangladesh extract (93930 tiles, 25544140 features). The comparison has to run with --threads 1: with several threads tilemaker's own output varies by a feature or two between runs of the same binary, because addObjectToSmallIndex() falls back to a per-thread pending list under lock contention.