mtmd: port gemma4uv/gemma4ua support — fixes Gemma 4 12B vision (#163)

[TheTom]

Summary

Fixes #163. The fork could not load the new Gemma 4 12B mmproj — it errored load_hparams: unknown projector type: gemma4uv because the fork lacked the gemma4uv (unified, encoder-less vision) and gemma4ua (unified audio) projectors. The SIGFPE in the original report is the upstream symptom; on this fork the model never got that far.

This cherry-picks the upstream fix ggml-org#24077 ("mtmd, model: allow skip build_vit()", commit a731805) into the fork.

Conflicts resolved

Clean 3-way except 3 non-mtmd files (converter + vocab). Notably: normalizer_lowercase lines that appeared in the diff context are pre-existing upstream code (from an earlier commit the fork hasn't synced), not part of ggml-org#24077 — dropped them so the port only carries ggml-org#24077's actual changes (suppress_tokens, the gemma4uv/ua graphs, projector enums, skip-build_vit).

Testing — verified end to end on CUDA (GB10, sm_121)

Downloaded ggml-org/gemma-4-12B-it-GGUF (Q4_K_M + mmproj-Q8_0) and ran llama-mtmd-cli on a real photo:

$ llama-mtmd-cli -m gemma-4-12B-it-Q4_K_M.gguf --mmproj mmproj-gemma-4-12B-it-Q8_0.gguf \
    --image cat.jpg -p "Describe this image in one sentence." --jinja
> A close-up, slightly blurry shot of a tabby cat's face with large dark eyes and a pink nose.

• mmproj loads (gemma4uv projector recognized, no SIGFPE)
• image encodes, model describes it accurately
• Builds clean on CUDA (sm_121) and Metal, 0 errors both
• Text-only path unaffected

Supersedes #166 (the standalone d_head guard is included here via ggml-org#24077's clip.cpp hunk).

Credit: @guarismo for the report, upstream ggml-org#24077.

[TheTom]

Updated — the branch now fully supports Gemma 4 12B vision on both backends.

Commit stack (6):

• mtmd, model: add Gemma 4 "unified" variant ggml-org/llama.cpp#24077 base gemma4uv/gemma4ua support
• fix(mtmd): floating point exception in new Gemma 4 12B model ggml-org/llama.cpp#24088 d_head/kq_scale FPE guard
• mtmd: enable non-causal vision for gemma 4 unified ggml-org/llama.cpp#24082 non-causal vision attention
• fix(mtmd): handle Gemma 4 audio projector embedding size ggml-org/llama.cpp#24091 audio projector embedding size
• Fix Gemma 4 Unified conversion ggml-org/llama.cpp#24118 converter fix
• ggml-metal: fall back to CPU for im2col when KH*KW exceeds threadgroup limit (new, fork-local)

Metal needed an extra fix. The gemma4uv patch-embedding conv decomposes to im2col, and the Metal im2col kernel launches KHKW threads/threadgroup. For Gemma 4's large patch conv, KHKW exceeds the Apple GPU 1024-thread cap, so it hit a runtime GGML_ASSERT(KH*KW <= max_threads) (this assert exists upstream too — no upstream fix). The guard makes Metal decline the oversized im2col so the scheduler runs that one op on CPU; rest of the graph stays on GPU.

Verified end-to-end on both:

• CUDA (GB10, sm_121): loads mmproj + describes a cat image correctly
• Metal (M5 Max): same, EXIT 0, accurate description (CLIP graph uses unsupported operators warning confirms the im2col→CPU split)

Builds clean on both backends, text-only path unaffected.