Reject software wgpu adapters to allow renderer fallback

[zao111222333]

This PR prevents iced_wgpu from selecting software-rendered wgpu adapters, such as llvmpipe (LLVM 17.0.2, 256 bits) on Red Hat 8.10, as the primary renderer.

Previously, when Backend::Best was used, the fallback compositor tried wgpu first. If wgpu successfully created an adapter backed by llvmpipe, iced treated it as a valid wgpu renderer and never fell back to tiny-skia.

The change now:

• Detects software adapters using wgpu::DeviceType::Cpu
• Also checks known software adapter names like llvmpipe, lavapipe, softpipe, and swiftshader
• Rejects these adapters before requesting a device in the window compositor
• Applies the same detection to the headless wgpu renderer
• Adds tests covering llvmpipe rejection and normal hardware adapter acceptance

[edwloef]

This feels like it should only happen if tiny-skia is actually enabled as a fallback, otherwise you just can't start up an iced app using software-rendered gpu backends.

[zao111222333]

This feels like it should only happen if tiny-skia is actually enabled as a fallback, otherwise you just can't start up an iced app using software-rendered gpu backends.

That's more reasonable, is there any "fallback or not" option in iced? So I can reuse that configuration

[pml68]

That's more reasonable, is there any "fallback or not" option in iced? So I can reuse that configuration

If tiny-skia and wgpu (or at least wgpu-bare) are both enabled then so is the fallback mechanism.

[zao111222333]

Now it only rejects software wgpu adapters when the tiny-skia fallback is available.