Keep the G-code identical to main on Clipper2

Since main moved to Clipper2, parts of this branch no longer gave the same
G-code as main:

- bridge_over_infill dropped expand(limiting_area, 0.3 * flow.spacing()) as
  a no-op. The offset is below one unit, but Clipper2 still unites its
  result, which splits and merges touching polygons and so changes the
  anchor lines. Running it on the polygons next to the bridge gives main's
  anchors without the whole-layer pass.
- tsp_remove_crossings stopped at the first repeated ordering, where main
  runs on to its pn * pn cap. The loop is periodic from that point, so it
  now takes only the steps to the ordering main stops on.
- With a single tile, the tiled booleans now make the plain call instead of
  cutting the clip to the tile first.

The tiled boolean test compared rings exactly. With the safety offset a tile
unites only the clip polygons near it, and Clipper2 can then round a
crossing 1 unit differently, so that case allows 1 unit.

Comments that named ClipperLib now say Clipper, and
docs/HLSD/polygon-clipping.md describes the tiled booleans.

G-code of the five handy models in four configurations and of a baked
texture relief is byte-identical to main. Colour-painted models still
differ: segmenting each island on its own splits a colour's region into
different pieces than one diagram over the layer, which on one model also
changes the first layer's tool order.

docs/HLSD/polygon-clipping.md

@@ -62,6 +62,23 @@ Clipper2 behaves differently by default, the wrapper adjusts it.
   clip do not leave slivers.
 - Open polylines are clipped with the non-zero rule and keep their direction.
 
+### Tiled booleans
+
+The sweep slows down with the number of edges crossing a scan line, so a layer
+cut into thousands of pieces makes every whole-layer boolean expensive.
+`diff_ex_by_piece()` and `intersection_ex_by_piece()` take a subject of
+non-overlapping `ExPolygons`, group them into tiles with
+`ClipperUtils::tile_expolygons()`, and run each tile in parallel against only
+the clip polygons near it, cut to the tile's box. Below 128 pieces there is a
+single tile, and they are the plain `diff_ex()` / `intersection_ex()`.
+
+The result covers the same area as the plain call. Without the safety offset
+the rings are the same. With it, each tile unites only the clip polygons near
+it, so a clip edge that the whole-layer union splits where it crosses a distant
+clip polygon stays whole, and a crossing with the subject can round 1 unit
+differently. The tiles' results are concatenated in tile order, so the order of
+the output `ExPolygons` differs from the plain call.
+
 ### Offsets
 
 - Before offsetting, input vertices closer than