Bump llama.cpp to dec5ca5 (b9763), release v0.8.28

[nyo16]

Summary

Updates the vendor/llama.cpp submodule from 845282461 to dec5ca557 (24 commits, tag b9763) and cuts release v0.8.28.

No NIF changes required. The only binding-relevant header touched is include/llama.h, and its sole functional change is one new accessor — llama_model_n_layer_nextn — added next to
llama_model_n_layer; the rest of the diff is whitespace realignment. The binding does not call the new function. All other headers the binding compiles against (ggml.h, ggml-backend.h,
chat.h, json-schema-to-grammar.h, sampling.h, speculative.h, common.h) are unchanged.

This range refactors the grammar generators behind the binding's json_schema_to_grammar_nif (common/peg AC parser #24869, until→GBNF refactor #24839, json-schema-to-grammar spacing #24835) even
though the header is unchanged — so both the JSON-schema and raw-GBNF smoke paths were re-verified.

See CHANGELOG.md for the full categorized commit breakdown.

Verification

• ✅ Full test suite: 158 passed, 4 skipped (7 smoke excluded)
• ✅ Smoke tests: 7 passed against real models — generation, streaming, chat template, JSON-schema grammar, raw GBNF (Llama-3.2-3B-Instruct), and embeddings (Qwen3-Embedding-0.6B)
• ✅ mix format --check-formatted clean
• ✅ Dialyzer: 0 errors

NIF rebuilt from source against the new submodule (LLAMA_BACKEND=auto, Metal).

Note: checksum.exs is intentionally untouched — CI regenerates it against the release artifacts after the tag is pushed.

You can also click the link GitHub printed on push: https://github.com/nyo16/llama_cpp_ex/pull/new/bump-llama-cpp-dec5ca5

---

Summary of what's done:

• Branch bump-llama-cpp-dec5ca5 pushed to origin
• Submodule 845282461 → dec5ca557 (24 commits, b9763)
• mix.exs @Version → 0.8.28
• CHANGELOG.md v0.8.28 entry added (categorized)
• checksum.exs left untouched (CI regenerates it)
• Verified: full suite + 7 smoke tests + format + Dialyzer all green