[AI] Fix disk cache load with dynamic encoder output shapes

[andriiryzhkov]

Problem

The disk cache for object mask encoder embeddings always fails on the first load after restarting darktable. The second load of the same image (within the same session) works fine.

Root cause

When the model loads with ORT_DISABLE_ALL (fallback for ORT versions where optimization crashes), encoder output shapes are reported as dynamic (-1). The cache file stores the resolved shapes from a previous inference (e.g. [1, 32, 256, 256]), but on load the validation compares them against the model's unresolved shapes ([-1, 32, -1, -1]) and rejects the mismatch.

After re-encoding, ctx->enc_shapes gets updated with resolved values, so subsequent cache loads in the same session succeed.

Fix

• Skip dimension validation when either the model or cache file reports a dynamic dim (<= 0)
• Copy resolved shapes from the cache file into ctx->enc_shapes after a successful load, so the decoder gets valid shapes instead of -1

[TurboGit]

Thanks!