Upgrade llama.cpp to b9106

[bernardladenthin]

Summary

Upgrade the pinned llama.cpp version from b9103 to b9106, incorporating improvements to Vulkan flash attention, CUDA argsort, and Mistral Medium 3.5 model conversion support.

Changes

• Vulkan flash attention refactor: The pipeline_flash_attn_f32_f16 pipeline changed from a per-type array of maps to a single map. Mixed K/V quantization types (e.g., Q4_0 K + F16 V) are now supported across all Vulkan FA paths (scalar, cm1, cm2) instead of only coopmat2. Per-type SPIR-V variants were replaced with two generic modules (flash_attn_f32_f16 and flash_attn_f32_f16_int8) that select K/V type at runtime via spec constants. A new flash_attn_dequant.glsl shader provides aliased SSBO views and unified dequantization logic.

• CUDA argsort enhancement: Added #include <cuda/iterator> to support CCCL ≥ 3.1 strided-iterator path.

• Mistral Medium 3.5 conversion support: Updated convert_hf_to_gguf.py to read "dim" key instead of "hidden_dim" for n_embd_text, and added resolution of negative img_break_tok_id placeholders from tekken.json or tokenizer.json.

Notes

All changes are internal backend improvements and conversion tool enhancements. No Java API changes are required.

https://claude.ai/code/session_013S1x9KhvBL6Pr3xuYoTtKh

[claude]

Review Summary

✅ Version bump executed correctly — All three required files updated per CLAUDE.md upgrade procedure:

• CMakeLists.txt: GIT_TAG b9103 → b9106
• README.md: badge and link updated
• CLAUDE.md: pinned version and breaking changes table updated

✅ Upstream changes properly documented — The three new rows in the breaking changes table (Vulkan flash attention refactor, CUDA argsort fix, Mistral Medium 3.5 mmproj support) accurately reflect that all changes are internal to GGML/CUDA backends or conversion tooling with no Java/JNI/server-API impact.

✅ No architectural changes — The b9103→b9106 delta touches only:

• Internal GPU backend optimizations (Vulkan, CUDA)
• Model conversion tooling (convert_hf_to_gguf.py)
  None of these affect the compiled upstream server sources or project's C++ headers.

Ready to merge — This is a low-risk version bump with proper documentation.