Problem

On Wayland with hybrid GPU laptops (e.g. NVIDIA + Intel), COSMIC panel applets consistently fall back to llvmpipe (CPU software rasterizer) instead of using the real GPU.

The root cause is a race condition during compositor initialization: the Wayland surface is not yet fully committed when wgpu probes adapter compatibility, causing VK_ERROR_SURFACE_LOST_KHR in vkGetPhysicalDeviceSurfaceCapabilitiesKHR. This falsely rejects real GPU adapters while llvmpipe survives (its Vulkan implementation doesn't perform the same surface validation).

This is particularly common with layer-shell panel applets, where the surface lifecycle differs from regular windows.

Observed behaviour

ERROR wgpu_hal::vulkan::adapter: get_physical_device_surface_capabilities: ERROR_SURFACE_LOST_KHR

Selected: AdapterInfo {
    name: "llvmpipe (LLVM 21.1.8, 256 bits)",
    device_type: Cpu,
    ...
}

Despite an NVIDIA T550 (discrete) and Intel Iris Xe (integrated) being available.

Fix

This PR contains two independent, incremental commits:

Commit 1: Trust DMA-BUF device preference without probing surface

When the Wayland compositor provides DMA-BUF feedback with vendor/device IDs that match an enumerated adapter, select it directly without the redundant is_surface_supported() probe. The DMA-BUF feedback already guarantees which device the compositor wants, and the surface will be (re-)created on the chosen adapter later anyway.

Commit 2: Retry adapter selection without surface when only CPU adapter found

When request_adapter() returns only a CPU software rasterizer, assume that real GPUs were falsely rejected by the stale surface and retry without the compatible_surface constraint. This lets wgpu's power-preference ranking pick the best real GPU available.

If the surface-less retry also yields nothing, the code falls back to whatever the original surface-constrained attempt returned (possibly the CPU adapter), so behaviour is never worse than before.

Edge cases considered

• Systems with only a CPU adapter (e.g. VMs without GPU passthrough): the retry produces the same CPU result and it is used as expected.
• Systems where DMA-BUF feedback is unavailable (e.g. layer-shell applets where get_wayland_device_ids returns None): Commit 1 is skipped, Commit 2 handles the fallback.
• Systems where the surface is valid: request_adapter returns a real GPU on the first try, the CPU check is false, and the existing code path is taken with no behavioural change.

Testing

Tested on a hybrid GPU laptop (NVIDIA T550 + Intel Iris Xe, driver 580.119.02, Mesa 25.3.5) running COSMIC desktop on Wayland. Before this fix, the applet always selected llvmpipe. After:

WARN  iced_wgpu::window::compositor: adapter selection: surface-compatible pick is
      llvmpipe (LLVM 21.1.8, 256 bits); retrying without surface constraint

Selected: AdapterInfo {
    name: "NVIDIA T550 Laptop GPU",
    device_type: DiscreteGpu,
    backend: Vulkan,
    ...
}