v2.9.1 — Captions placed in near-linear time

Charts with many edge captions lay out far faster, with every caption where
it was. Output for every chart is byte-identical to 2.9.0.

Fixed

• Charts with many edge captions lay out far faster. The caption placer
  scored every candidate position against every node, caption, bend,
  crossing and arrowhead in the chart, so its cost grew about n^2.5 with the
  number of labeled edges: a fan of 320 labeled edges took 18 s (debug
  build). It now reads only what a spatial grid finds near each candidate,
  skips a candidate whose unavoidable penalties already lose to the best one
  so far, counts the bends and crossings in a caption's way without sorting
  them, and tests only route segments that could cross: the same fan takes
  0.3 s, 480 labeled edges 0.5 s, and 160 labeled self-loops 0.07 s instead
  of 4.2 s. The placement is bit for bit what it was: every skipped term was
  exactly zero and every skipped candidate could not have won, and a test
  holds the new placer to the old one over real layouts and 300 seeded random
  ones. (#97, #99, #101)