fix(mtmd): guard clip d_head/kq_scale against n_head==0 (Gemma 4 12B SIGFPE)

[TheTom]

Summary

Fixes #163. Loading the Gemma 4 12B mmproj crashes with SIGFPE, Arithmetic exception at clip.cpp:250.

Root cause: Gemma 4 12B's vision tower has no attention encoder, so hparams.n_head == 0. The clip_graph constructor's init list does:

d_head(n_embd / n_head),                  // integer divide-by-zero -> SIGFPE
...
kq_scale(1.0f / sqrtf((float)d_head)),    // 1/sqrt(0) once d_head is fixed

Fix

Guard both, identical to the upstream fix for ggml-org#24085 (closed):

d_head(n_head > 0 ? n_embd / n_head : 0),
kq_scale(d_head > 0 ? 1.0f / sqrtf((float)d_head) : 0.0f),

When the tower has no encoder, d_head/kq_scale are unused, so 0 is safe.

Testing

• mtmd target builds clean (Metal, clip.cpp.o compiles).
• Matches the canonical upstream patch already merged in ggml-org/llama.cpp.
• Did not repro locally (no Gemma 4 12B mmproj on hand); the guard is a strict superset of the crashing path.

Thanks to @guarismo for the clear report and backtrace.

[TheTom]

Superseded by #168. Testing on the actual gemma-4-12B revealed the fork was missing the whole gemma4uv projector (it errored unknown projector type: gemma4uv before ever reaching the d_head division). #168 ports the full upstream fix (ggml ggml-org#24077), which includes this d_head guard, and is verified end-to-end (loads the mmproj + describes an image correctly). Closing this in favor of that.