Fix inscribed circle initialization

[unjambonakap]

Fix a bug in MaximumInscribedCircle that gives a bad result

import shapely, shapely.constructive
import numpy as np
poly = shapely.from_geojson('{"type":"Polygon","coordinates":[[[0.0,-10.0],[-7.07107,-7.07107],[-10.0,0.0],[-7.07107,7.07107],[0.0,10.0],[7.07107,7.07107],[10.0,0.0],[7.07107,-7.07107],[0.0,-10.0]]]}')

u = shapely.constructive.maximum_inscribed_circle(poly, tolerance=0.1)
a, b = np.array(u.coords)
print(a,b,np.linalg.norm(a-b))

Actual:
[0. 3.535535] [-2.28553181 9.05330273] 5.972387826209451

Expected:
[0. 0.] [ 8.53553655 -3.53553126] 9.238796754579768

[pramsey]

As you can see, this fails regression, though only regression tests on invalid inputs. Unfortunate!

• in test 5, the result moves from the middle of the (invalid) geometry to the end, which is "less good" though again we are talking about invalid input and a zero-radius output...
• similar story in test 7, moving from the middle to the end

---> group: geos::algorithm::construct::MaximumInscribedCircle, test: test<5>
     problem: assertion failed
     failed assertion: `x coordinate does not match: expected `1.5000000000000000e+02` actual `1.0000000000000000e+02` with precision `1.0000000000000000e-02``

---> group: geos::algorithm::construct::MaximumInscribedCircle, test: test<7>
     problem: assertion failed
     failed assertion: `x coordinate does not match: expected `2.0000000000000000e+00` actual `1.0000000000000000e+00` with precision `1.0000000000000000e-02``

[pramsey]

The JTS tests seems to also expect the result to be the same as the centroid point, so I am not at all clear on why when you initialize to the centroid you then get back a non-centroid answer. That seems like something we should understand before just changing the regression answers.

[unjambonakap]

The fact that it used to hit the centroid is purely by luck:
Before test5:

• First cell distance is bad (because computed with (0,0), outside)
• first iteration, gets improved with 0 (this happens to be the centroid)
• no more improvements

Now:

• first cell is the best one, attributed to the computed interior point (on the boundary, 100 100)(