[AI] Speed up AI inference in neural restore module

[andriiryzhkov]

Summary

• Specialize dynamic ONNX model sessions to a concrete tile size at load time using AddFreeDimensionOverrideByName, so JIT-compiling execution providers (MIGraphX, CoreML, TensorRT) compile the kernel once per session instead of recompiling on every inference call
• Cache the working tile size in darktablerc (keyed by model/scale/provider) so subsequent runs skip OOM retry entirely
• Persist AMD GPU caches (MIOpen kernel DB + MIGraphX compiled programs) and Intel OpenVINO compiled blobs to the user cache directory, eliminating multi-minute recompilation on every launch
• Add spatial_dims field to model config.json so each model declares its symbolic dimension names

[TurboGit]

Works for me, thanks!