[AI] Fix corrupted AI output when provider is set to auto

[andriiryzhkov]

A user on pixls.us reported posterized blue/magenta raw denoise output on macOS – but only with "auto" selected. Explicit "CoreML" or "CPU" worked fine.

The rawdenoise-nind manifest declares cpu_only: { model_bayer: [coreml] }, telling the backend that the Bayer head needs CoreML's CPU compute units (the ANE/GPU path produces garbage). That check ran for explicit "CoreML" but was exempted for "auto" – so on macOS "auto" enabled CoreML without USE_CPU_ONLY and ran on the ANE.

Rather than special-casing macOS, AUTO and cpu_only now flow through one resolver in backend_common.c:

• AUTO walks a per-platform candidate list (macOS: [CoreML], Windows: [DirectML], Linux: [CUDA, MIGraphX, OpenVINO]) and applies cpu_only to each. First candidate that's available and either not banned, or (CoreML) tolerable via USE_CPU_ONLY, wins.
• _enable_acceleration no longer carries an AUTO case – the resolver guarantees a concrete EP.
• Single-candidate platforms skip dt_ai_probe_provider; probe results are memoized per process.

Linux AUTO on a model that bans one GPU EP now falls through to the next instead of dropping straight to CPU. Explicit-EP and CPU paths are unchanged.

[TurboGit]

Thanks!