Robust circumcenters when the hull is collinear

[Fil]

The failure reported in #141 came from the fact that the first point of the first triangle belonged to the collinear part of the hull, thus making the circumcenters of degenerate triangles (which need to be "projected to the infinite") go to either side arbitrarily. We want them to go "outwards".

The solution suggested here is to make sure we use a non-collinear reference (which I took to be the barycenter of the hull—could have been something else; but the first point of the first triangle was clearly a wrong bet).

I've reduced the test case to:
points = [[10, 190]].concat(Array.from({length: 7}, (d, i) => [i * 80, (i * 50) / 7]))

before

after

fixes #141

[mbostock]

This looks like an improvement. 👍 Is there still going to be an issue when all the points are collinear, since in that case the barycenter of the hull will also be collinear?

[Fil]

The case where n>2 points are all aligned is addressed differently: we detect this case in

d3-delaunay/src/delaunay.js

Line 55 in 48d091f

// check for collinear

and jitter the points.