[arrayTools] fix sectRect boundary case

[anthrotype]

use '>' instead of '>=' when comparing intersection of two rectangles, so we don't skip boundary case when the two touch

Fixes #3352

[justvanrossum]

I notice I've done the same change in my JS port in Fontra, so that makes me +1.

[anthrotype]

@simoncozens I noticed that when making this change to sectRect, from no intersections, we now find 4 intersections for the test curves (from dad5bc6) that touch exactly at t=0.5, but they look like duplicates.

[Intersection(pt=(49.853515811264515, 74.99971389770508), t1=0.4990234375, t2=0.4990234375),
 Intersection(pt=(50.146484188735485, 74.99971389770508), t1=0.5009765625, t2=0.4990234375),
 Intersection(pt=(49.853515811264515, 74.99971389770508), t1=0.4990234375, t2=0.5009765625),
 Intersection(pt=(50.146484188735485, 74.99971389770508), t1=0.5009765625, t2=0.5009765625)]

I see that _curve_curve_intersections_t function has a precision parameter set to 1e-3, which is used for two slightly different things:

1. when determining if an intersection was found, by comparing it against the area of the bounds rectangle

    if rectArea(bounds1) < precision and rectArea(bounds2) < precision:
        return [(midpoint(range1), midpoint(range2))]

2. when rounding the ts in order to prune duplicates:

    unique_key = lambda ts: (int(ts[0] / precision), int(ts[1] / precision))
    seen = set()
    unique_values = []

    for ts in found:
        key = unique_key(ts)
        if key in seen:
            continue
        seen.add(key)
        unique_values.append(ts)

Maybe we need to use precision*precision (squared) when checking the areas (given area itself is not a single-dimension dinstance but... an area), so that we then obtain more precise intersections whose ts will get rounded off with given precision and will become equal and deduped?

[anthrotype]

basically this bfd874a

[simoncozens]

This stuff (any kind of cubic-cubic intersection logic) is horribly complicated and full of floating-point imprecision and edge cases and I've basically hit the limit of my knowledge here. (Actually I hit the limit of my knowledge here.) Raph has been working on a new algorithm for Kurbo but I'm not sure what state that's in.