PDF tiling patterns: colour-through-base fix, BBox clip, registry dedup

[andiwand]

🤖 Generated with Claude Code

Follow-ups on the stage-4.10 tiling-pattern work (#574), addressing the Codex review comments plus some cleanup.

Correctness

• Uncoloured-pattern colour through the base space. ColorSpaceDef::to_rgb returned black for ColorSpaceKind::pattern, ignoring its base. When a /PaintType 2 pattern is selected through a named Pattern colour space (/CS1 [/Pattern /DeviceRGB]), set_color feeds the leading components into to_rgb, so /CS1 cs 1 0 0 /P scn painted black instead of red. to_rgb now defers the pattern case to its underlying base. Adds a regression test (the existing test used the inline /Pattern cs form, where color.def stays null and the device-by-count path masks the bug).
• Clip each tiling-pattern cell to its /BBox (ISO 32000-1 8.7.3.1). When /XStep//YStep differ from /BBox, marks drawn outside the cell — or in the gap when a step exceeds the BBox — leaked into the repeated tile. The tile content is now wrapped in a <clipPath> rect sized by the BBox. (An overlapping lattice, a step smaller than the BBox, can't be expressed as a single SVG <pattern> and is left unreproduced.)

Cleanup / dedup

• Extract a shared DefsRegistry base for the three per-page <defs> registries (clip/gradient/pattern), which carried identical m_page/m_count/m_id_by_signature/m_defs members, an identical defs(), and the same intern-and-mint-id idiom.
• Extract an svg_matrix helper for the repeated matrix(a,b,c,d,round2(e),round2(f)) serialization (gradient/pattern transforms and the image fragment).
• Use has_value for the pattern /Resources lookup, matching the /BBox check.

Tests

All 200 PDF tests pass locally (199 + the new scn_uncoloured_tiling_pattern_colour_through_base).