[jwinarske]

This is part of moving the parity baseline to drm-cxx 2a050f5. It ports the allocator half of upstream's 58cf1a7 (drm-cxx#363). The parsing half is #124.

drmkit enforced a CRTC's blend-stage limit (P-46) inside the search. Tests over the limit were refused without asking the kernel, and the backtrack dropped layers until the assignment fit. Three things differed from the reference:

• Which layers survive: the ones the score-ordered greedy pass happened to place. A high-priority layer that pass left out stayed composited while low-priority ones held planes. That's the symptom #363 fixed.
• Pinned and externally bound layers: they weren't counted against the limit.
• The canvas's stage: it was counted only while a canvas plane was armed.

The new cap_to_blend_limit runs before the full search, as the reference's does:

• pinned and externally bound layers hold stages of their own;
• one stage is left for the canvas when anything composites;
• the rest go by keep_priority, ties in the caller's order;
• layers past the budget are composited.

The in-search check stays as a backstop.

Tests:

• New: the_blend_limit_keeps_the_layers_most_worth_a_plane. With a limit of 3 and the high-priority layers at both ends of a five-layer stack, it now places exactly those two. Without the cap it placed 1, 4 and 5.
• no_test_arms_more_planes_than_the_crtc_can_stack now expects the reference's answer: two placed, three composited, one stage for the canvas. It expected three placed before.

Docs: P-105 row and a CHANGELOG entry.

Checks: clippy over the workspace with all features, workspace tests, the scene, planes and scene-set tests on vkms, and rustdoc all pass. vkms has no blend limit, so nothing changes there.