[Fi3w0]

0.90.200 adds the idle cleanup_texture_cache call site that plan §3 deferred until after the Smithay migration.

Why: Smithay frees dropped GL textures, EGL images and buffers only after a rendered frame (GlesRenderer::cleanup on unbind, or a no-damage render_frame), and only for the renderer that drew. That leaves two gaps:

• Anything released after an idle desktop's last frame waits for the next frame.
• Render-only GPU nodes, which hold cross-GPU imports, are never swept, because ImportBridge::cleanup_texture_cache only reaches the primary renderer.

Change: the udev maintenance timer now calls GpuManager::cleanup_texture_cache (upstream 347b2b30).

• Rate-limited to once every 5 s, never per frame or per output.
• Skipped while the DRM device is paused (VT switched away).
• Takes the manager with try_borrow_mut, so it can never collide with a render.
• Upstream 79bbed5e makes a context current only when its destruction queue is non-empty, so an idle node isn't woken from runtime suspend (Smithay #2178/#2180).

Validation:

• Formatting, 708 locked all-feature/all-target serial tests, strict Clippy and the locked release build pass.
• Tested on a native install of .200, NVIDIA only (AMD excluded by config).
  • Idle for 2 min: flat at 75 MiB GPU, PSS about 100 MB.
  • Three rounds of 7–8 kitty windows on a hidden workspace, killed while hidden: GPU and PSS went back to baseline, and fds went 264 → 187.
  • 200 grim captures: no errors.
• Not run: VT switch, DPMS, OBS with the sweep active.

Separate finding, not caused by this change: the first capture of each output grows the NVIDIA driver's /dev/nvidiactl mapping (+44 MiB on 2560×1440), and it stays at that level. It's bounded, and the sweep can't release it. Tracked separately.