#26 fix(drm): allocate hardware cursor images as KMS dumb buffers

Fixes neferwl#22.

Problem

On nvidia-drm, KMS refuses the hardware cursor plane with EINVAL (cursor plane refused; hardware cursor off). NeferWL then uses the software cursor, so every pointer motion forces a full frame. The cursor images were linear images that Vulkan allocated and exported, the same allocate-in-Vulkan, scan-out-in-KMS path that corrupted the primary plane on NVIDIA.

Fix

• Allocation: cursor images are KMS dumb buffers (CREATE_DUMB, MAP_DUMB and mmap, AddFB2 by GEM handle), written by the CPU when the cursor image changes. Moves are unchanged: they only update the plane position.
• Fallback: dumb buffers, then Vulkan cursor images, then the software cursor. Vulkan images are tried both when dumb allocation fails and when the plane refuses the dumb images in the TEST_ONLY probe; the current image is rewritten into them.
• Log: hardware cursor logs alloc (dumb or vulkan, a warning for vulkan) and the size.
• HDR cursor conversion: moved to internal/adapters/hdrcursor, so the drm adapter and the renderer write the same pixels. A recorded vector checks byte-identical output.

Verified

• make check, make race, make mocks-check and staticcheck ./... pass.
• Real-device dumb-buffer test (allocate, map, write, AddFB, read back; no modeset or commit) passes on amdgpu and on nvidia-drm 615.78.08.
• Unit tests with generated mocks cover the allocator order, the TEST_ONLY refusal and swap (in-flight state, queued and hidden images), writes, crop and pitch, and cleanup.

Not verified

• NVIDIA cursor plane acceptance with a display attached. This needs a test on an NVIDIA machine: the log should show "message":"hardware cursor" with "alloc":"dumb" and no cursor plane refused, and the cursor should show at startup.