Reduce Metal backend memory use

[elcritch]

Summary

• Allocate backdrop textures and CPU rect-mask arrays only when needed. Release backdrop textures after three consecutive frames without blur, and release mismatched optional textures on resize.
• Drop the CPU copy of static indices after uploading them to Metal. Read screenshots directly into the Pixie image buffer and swap BGRA channels in place.
• Add a Metal regression test and a reproducible memory benchmark.

Memory profile

Release ARC build on Apple M3 Pro, macOS 15.7.9. Three runs per scenario; values are medians. Workload: 1024×768 frame, 1024² atlas, 1024 max quads.

Scenario	Before	After
Plain frame, Metal allocated memory	11.28 MiB	8.25 MiB
Three frames after blur stops, Metal allocated memory	14.31 MiB	8.25 MiB
Full screenshot, peak process RSS	22.66 MiB	19.47 MiB

An active blur frame remains at 14.31 MiB in both versions. The benchmark uses MTLDevice.currentAllocatedSize and /usr/bin/time -l; its source is in tests/bench_metal_memory.nim.

Tests

• atlas-run tests — 46/46 passed.
• Metal regression covers lazy allocation, idle release, resize/recreation, and screenshot channel order.

Threading and GC

No thread or GC configuration changes. The three-frame idle window matches the existing maximum number of in-flight Metal frames.