Fix rounding bug and improve arc discretization

[gpeairs]

Fix #24. Also use the improved circular_arc method to render Turns with circular arcs—since the curvature is constant and known, we can easily discretize to a tolerance using a simple formula without resorting to adaptive methods.

Some notes:

• The relationship between adapted_grid and tolerance in the sense of atol's "max distance between polygon and exact edge" is complicated. One might guess that grid_step = 1.0μm, max_change=5° means that we start with a grid step of 1 micron (true) and refine until the max change in direction between steps of the underlying curve falls below 5 degrees (we get much smaller change depending on the radius but it's hard to follow why, and I'm not sure whether or not that's a bug). In any case, for tight curves, the distance from the true curve can be significantly more than 1nm, and for large radius of curvature, far more points are used than are necessary for 1nm tolerance (1μm steps give 1nm tolerance at 125μm radius, and tighter tolerance for larger radii).
• This discretizes the circular arcs along a styled Turn with the same angular step, such that the curve with the largest radius meets atol (and thus so do all curves with smaller radii). An alternative implementation could discretize the inside edge of a trace with fewer points than the outside edge. There's no official promise that the curves have the same number of points, but I think it's nice that the width of the discretized trace is constant. All other curves are discretized according to the Turn without regard to the style, so they also have the same number of points on each curve. Eventually, we may want to simplify path rendering by first converting to CurvilinearPolygons using OffsetSegments, then discretizing the offset segments. In that case, we could preserve this behavior by going back to choosing the discretization according to the underlying Turn curve, although then tolerance might not be met on outer offset curves.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

📢 Thoughts on this report? Let us know!

[gpeairs]

Even without rigorous benchmarking, we can see that this is a significant speedup over using adapted_grid for Turns. On DemoQPU17, we previously had (first execution, subsequent execution):

Rendering to polygons: 5.758117 seconds (10.51 M allocations: 615.602 MiB, 8.76% gc time, 88.12% compilation time)
Rendering to polygons: 0.654979 seconds (5.79 M allocations: 299.579 MiB, 4.17% gc time)

This PR takes off ~90% of the rendering time after compilation:

Rendering to polygons: 3.857135 seconds (4.37 M allocations: 317.261 MiB, 1.42% gc time, 97.37% compilation time)
Rendering to polygons: 0.077314 seconds (502.31 k allocations: 57.515 MiB, 5.17% gc time)