Bump llama.cpp to b9826, release v0.8.30

[nyo16]

Summary

Bumps the vendor/llama.cpp submodule from b3ce5cedf (b9789) to 9bebfcb4b (b9826) — 37 commits — and cuts release v0.8.30.

No NIF changes required. All headers the binding compiles against are unchanged except common/common.h, whose diff is server/CLI-only (new LLAMA_EXAMPLE_DOWNLOAD enum, const on
common_params_model::get_name() + new empty() helper, has_dft() refactor, removed server-side skip_download flag) — none consumed by the binding. The range also resyncs ggml to 0.15.3 underneath the
statically-linked libraries, so the NIF was rebuilt clean.

Verification

• mix compile --warnings-as-errors — clean
• mix test (Qwen3.5-0.8B + Qwen3-Embedding-0.6B from ~/Downloads) — 216 passed, 1 skipped
• mix test test/smoke_test.exs --include smoke — 7 passed (generation, streaming, chat templates, JSON-schema grammar, raw GBNF, embeddings)
• mix format --check-formatted — clean
• mix dialyzer — 0 errors