Fall back to finite-diff tangent when analytic derivative is zero

[andriiryzhkov]

Fixes #21525 and #21557

Path masks with ctrl == corner at a segment endpoint (potrace routinely emits this) hit the exact-zero branch in _path_border_get_XY, which returned DT_INVALID_COORDINATE. Gap-fill then bailed on invalid input and the border polyline connected surviving points across huge distances - the chaotic wide crossing lines reported.

Fall back to a finite-difference sample of the curve for those cases. Non-degenerate paths hit the analytic branch as before, so existing masks and the integration tests are byte-identical.

Replaces the earlier attempt in be7259b (which switched to finite-diff unconditionally and shifted every path mask by a few pixels)

[andriiryzhkov]

For release notes:

In Bug Fixes:

• Fix feather rendering on path masks whose control points coincide with their corners; previously produced wildly displaced border lines instead of a proper offset outline

[TurboGit]

Thanks!