[asimage] Make the two ellipse fill spans mirror each other

[tekinertekin]

This Pull request:

The function asim_ellips2() creates a tilted ellipse using a scanning technique, rather than through flood fill. On each step of the walk, two spans are created, one above and one below the center, based on the rotational symmetry of the shape. Thus, these spans should have been reflections around the center, but they weren't. While the upper span terminated at x2 - 1, the lower span began at x - x2 - 1, which mirrors to x2 + 1.

Changes or fixes:

Two columns remain unpainted inside the shape on the upper rows, while two columns outside it become painted on the lower rows. The gap is apparent for every tilt angle and increases with the angle; at 45 degrees for rx=21, ry=9 there are eight unpainted pixels inside the shape, including the one that is mentioned in the report. The straight cases are not affected since 0, 90, 180, and 270 degrees exit early in asim_straight_ellips().

Using an explicit name for the span end exposes the symmetry and allows each walk branch to set the end from its own parameters. Only one of them needs a different end: it draws the anti-aliased outline at x2+2, so the fill can go to x2+1. The other two branches retain the previous end, so the output of theirs is unchanged.

Measured against the interior of the shape itself using its outline, over 610 combinations of angle and radii: number of interior pixels that were not painted decreases from 7698 to 30, number of exterior pixels that were painted decreases from 7496 to 1102; neither of the combinations gets worse on any of these criteria. Unfilled ellipses and straight-ellipse path paint identically before and after the change, for every tested angle and radius.

Before and after

DrawEllips2(32, 32, 21, 9, 45, "#FF2277CC", -1), the reproducer from the issue. The red cells in the magnified crop are the pixels the fill leaves empty; that region is scaled 22x with a pixel grid.

	master	with this change
full 64x64
defect region, magnified

The gap is not specific to 45 degrees. At 85 degrees the same radii leave 30 pixels empty instead of 8:

	master	with this change
angle 85

The tests already exist

graf2d/asimage/test/tasimage_ellipse_draw.cxx, added in #23119, already covers this, including the assertion for the pixel named in the issue:

test	on master	with this change
TASImage.FilledEllipsOpaque	fails	passes
TASImage.FilledEllipsHighAlpha	fails	passes
TASImage.FilledEllipsSemiTransparent	fails	passes
TASImage.FilleEllipsLowAlpha	fails	passes

All four fail on the same assertion, the one added for this bug:

unsigned x = kSize / 2 + kSize / 5 + 1;   // 45
unsigned y = kSize / 2 - kSize / 5 - 2;   // 18
EXPECT_NE(argb[y * kSize + x], centreBefore) << "the DrawEllips2 does not fill point inside the " << name;

The corner, centre and fill-count assertions pass both before and after; only that one flips. The fill count also moves closer to the expected area, 699 to 689 against an expected 691, rather than further from it.

I could not run the gtest binary itself, since this build has asimage=OFF, so I reimplemented the four assertions against asim_ellips2() directly in a standalone harness over the vendored draw.c. CI will be the real check.

Assisted-by: Claude (Anthropic); the change was AI-assisted, then reviewed, measured and verified by the author.

Checklist:

• tested changes locally
• updated the docs (if necessary)

This PR fixes #23120