server: identical multimodal embeddings differ with default unified KV cache #42

Upstream: ggml-org#28656

Two identical multimodal inputs in one /embeddings request return different vectors under the server's default unified KV cache.

Environment: Linux aarch64, NVIDIA GB10 / CUDA 13, current main 434ddbbc0e30522e897670681e503b797c12b7c1; ggml-org/tinygemma3-GGUF Q8_0 plus its mmproj and the repo's tools/mtmd/tests/test-1-positive.png.

LLAMA_MEDIA_MARKER='<__media__>' llama-server \
  -m tinygemma3-Q8_0.gguf --mmproj mmproj-tinygemma3.gguf \
  -ngl 99 --embeddings -c 4096 -b 512 --seed 42 --no-jinja

Send one POST /embeddings whose content contains the same item twice:

{"prompt_string":"What is this: <__media__>\n","multimodal_data":["<same base64 PNG>"]}

Observed: default kv_unified=true gives max|delta|=0.0705835, cosine 0.99989352. Adding -np 4 selects kv_unified=false and both vectors are bit-identical.

This isolates the failure to CUDA multimodal embeddings with the unified KV path; text-only and CPU are reported unaffected. Introducing commit not identified.