Merge tied, nearly parallel plain edges in their exact order in no_fit_polygon

[fontanf]

Fixes the cause of the failures remaining in fontanf/packingsolver#598 with the latest shape (after #71 and #72).

Problem

The convex no_fit_polygon merges the edges of both shapes in angular order. It compares directions with direction_strictly_lesser, whose tolerance is meant for the floating-point noise of arc tangents. But, on unit vectors, it also ties plain edges up to ~1e-6 rad apart. Tied plain edges were then always merged fixed first, which is the wrong order whenever the orbiting edge has the smaller direction: the vertex between them is replaced by one inside the actual Minkowski sum.

In packingsolver#598, instance B's hexagon has two edges ~9.5e-7 rad from parallel (edges 0 and 3). In its self-NFP, this happened on one side only:

• the vertex (58.8411, -157.0224) was replaced by (130.8797, -104.8274), one edge length away;
• the NFP was off the exact Minkowski sum by up to ~7e-5 (area 112131.8571 instead of 112131.8683), and was not centrally symmetric anymore.

The periodic packing then unions the inflated NFPs of a shape and of its half-turn. These should be identical, but differed by a few micro-units, and compute_union failed with "face area is not positive" or "outline area is not positive", or overflowed the stack.

Fix

Order tied plain edges by their exact directions (strictly_lesser_angle), and only merge the fixed one first if they are exactly parallel. Ties involving arcs are unchanged.

Testing

• New NoFitPolygonConvexTest case: the self-NFP of that hexagon, against the exact Minkowski sum (computed with rational arithmetic). It fails on main and passes with the fix. Its grid-sampling oracle check takes ~4 s.
• Issue558SawtoothSelfPairing: both expected_outputs variants are re-recorded, without and with fused multiply-adds. The result keeps the same structure (1 component, 883 elements, no self-intersection), with ~1.1e-4 more area, as expected since the old order put some vertices inside the actual Minkowski sums. Before re-recording them, I checked that a plain x86-64 build reproduces the old variant 0 exactly, and an x86-64 build with -mfma the old variant 1 (the AArch64 one).
• All 497 tests pass, in both a plain and an -mfma build.
• With packingsolver built against this branch, instances A, B and C of packingsolver#598 and 19 random perturbations of each (vertices moved by up to 1e-5) all solve: 60/60, against 42/60 with main.