fix: validate vision projector output dim against LLM hidden size

[fszontagh]

Summary

Pairing an mmproj with an LLM of a different size crashed with an opaque GGML_ASSERT(a->ne[d] == b->ne[d]) inside ggml_concat, from LLM::splice_image_embeds splicing image embeds between text embeds of a different width.

The projector output dim is already detected from the weights, so compare it against the LLM hidden size at load time and disable vision with a message naming the cause, matching the existing "no vision weights detected" fallback. The comparison only runs when the dim actually came from a merger tensor, so a checkpoint whose naming is not recognised keeps its previous behaviour instead of being rejected by the default.

Related Issue / Discussion

None.

Additional Information

Reproduced with Qwen3VL-4B-Instruct-Q4_K_M.gguf (hidden size 2560) and an mmproj-F16.gguf whose GGUF header is general.name = Qwen3-Vl-8B-Instruct, clip.vision.projection_dim = 4096. Before the change it aborts in ggml_concat; after it logs

vision projector output size (4096) does not match LLM hidden size (2560), ... vision disabled

and completes text-only instead of aborting. A correctly paired checkpoint is unaffected: verified with a Qwen3-VL encoder carrying its own vision tower, which still enables vision and produces a normal edit.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.