feat: Update vendored llama.cpp to include ModernBERT support

[clemlesne]

Context

The current release (v0.3.16, Aug 15 2025) vendors llama.cpp at commit 4227c9b (Aug 14, 2025). ModernBERT architecture support was added to llama.cpp on Dec 22, 2025 in PR ggml-org/llama.cpp#15641 (src/models/modern-bert.cpp). The vendor submodule hasn't been updated in ~7 months.

Problem

Loading a ModernBERT GGUF model fails with:

llama_model_load: error loading model: error loading model architecture: unknown model architecture: 'modernbert'

The binary only knows: bert, falcon, gemma, gpt2, llama, mamba, qwen, rwkv, stablelm — no modernbert.

Attempted fix

We tried updating the vendor submodule to llama.cpp b8373 (Mar 16, 2026) and building from source. Results:

1. CMake fails with LLAVA_BUILD=ON (default) — vendor/llama.cpp/tools/mtmd/CMakeLists.txt:37 has a set_target_properties error incompatible with the binding's CMakeLists.txt
2. CMake succeeds with -DLLAVA_BUILD=OFF — builds and installs fine
3. Runtime ABI mismatch — the Python bindings call llama_get_kv_self which no longer exists in b8373 (symbol not found). The C API has diverged significantly in 7 months.

So it's not just a submodule bump — the Python ctypes bindings need updates to match the current llama.cpp C API.

Why it matters

Several recent high-quality embedding models use ModernBERT architecture:

• granite-embedding-small-english-r2 (IBM, 47M params, 384d, BEIR NDCG@10 50.9, Apache 2.0)
• modernbert-embed-base (Nomic, 149M params, Apache 2.0)
• gte-modernbert-base (Alibaba, 149M params)

These models can't be used with llama-cpp-python until the binding is updated.

Request

Update the Python ctypes bindings and CMakeLists.txt to be compatible with a recent llama.cpp build that includes ModernBERT support (post Dec 22, 2025).