Contour maps optimization - #1849

The test that decides whether a projected segment can be kept as it is
was adapted from straightAndDomain() in lib/cartopy/trace.pyx of Cartopy
0.25.0: the position along the segment and the separation of the
projected midpoint from it, the tolerance that narrows towards the ends
of the segment, and the requirement that the map cover the segment.

Cartopy is released under the BSD 3-Clause License, which requires that
redistributed source keep its copyright notice, conditions and
disclaimer. The notice is reproduced verbatim above _project_lines,
where it travels with the source into every distribution, and the
docstring and comments name where the test comes from.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

On an orthographic map about half the lines of a field are on the far
side of the globe, and none of their points can be projected. They
were still given to Cartopy one at a time, only to come back empty.

For a segment whose ends both cannot be projected, Cartopy's bisection
in lib/cartopy/trace.pyx accepts the end at once and _project_segment()
adds nothing, without looking at the points in between. A line with no
point that can be projected is therefore never on the map, and is now
dropped from the points already projected. The result still matches
Cartopy line by line for 10 projections on four datasets and both
methods, and projecting the lines of a 197k-face noisy field on an
orthographic map takes 0.85 s instead of 1.9 s.

The rule is credited to Cartopy next to the segment test, under the
same BSD 3-Clause notice.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…tour

The "interpolated" method handed its triangulation to Matplotlib's
tricontour, which spent ~2.2 s of its ~2.6 s on a 786k-face mesh
setting up the neighbours of every triangle before contouring.

_crossed_edges, compiled with numba (cache=True, nogil=True), finds the
two edges where a level crosses each triangle that it crosses. Each
crossing point is known by its edge, which the triangles on either side
share, so the pieces are joined into lines by the existing
_join_segments kernel, and the points are interpolated along their
edges all at once. The triangulation is unchanged.

The lines are the same segments as tricontour's, to within 1e-9, on
eight datasets including node-centered data and faces with repeated
nodes; tricontour also left a few slivers ~1e-14 long on the
antimeridian, which are gone. A closed loop may start at a different
point, and a loop that crosses the antimeridian is then cut there
instead, which draws the same. Contouring a noisy field on the 786k-face
mesh takes 0.68 s instead of 3.4 s, of which 0.3 s is the
triangulation.

The splitting of lines at the antimeridian, shared with the edges
method, moves to _split_lines_at_antimeridian.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

contour-maps frames the lines by the area of the grid, set as their
extents, so that they never widen a shaded plot they are overlaid on.
Since the lines are projected before GeoViews sees them, HoloViews
framed the lines and their labels by the projected line data instead,
which reach closer to the antimeridian than the shaded faces: a Robinson
map of a field whose lines reach the edge of the map was 34 km wider on
each side than the shaded plot.

The projected area of the grid is now also the default xlim and ylim of
the lines, which HoloViews keeps to; options passed to .plot.contour()
still win. The frame then equals the shaded plot's own on Robinson,
Robinson centred on 180, orthographic, Lambert conformal and
PlateCarree maps, alone and overlaid, with the frame sizes and label
positions of contour-maps.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>