Bump llama.cpp to b3ce5ce (b9789), release v0.8.29

[nyo16]

Update vendor/llama.cpp submodule from dec5ca557 to b3ce5cedf (26 commits, tag b9789) and cut release v0.8.29.

NIF changes

None. include/llama.h, the ggml headers, and common/{json-schema-to-grammar,sampling,speculative}.h are all unchanged. The two binding-compiled headers that did change touch only server-side code:

• common/common.h — bumps a server default (checkpoint_min_step 256 → 8192).
• common/chat.h — reworks the chat message-splitting API (common_chat_msg_span.role becomes a common_chat_role enum; new common_chat_msg_spans / common_chat_msg_delimiters containers replace
  the free-standing common_chat_split_by_role; common_chat_params.message_spans → message_delimiters). All consumed by the server, not the binding.

The NIF only calls common_chat_templates_inputs, common_chat_msg, and common_chat_templates_apply. The implementation behind that last call (common/chat.cpp) was refactored as part of the same
work, so the chat-template smoke path was re-exercised against the freshly built NIF.

Notable upstream changes

• model: LFM2.5-ColBERT-350M, LFM2.5-Embedding-350M (#24913); Granite Speech Plus (#24818)
• quant: fix quantizing MoE with MTP (#24986)
• server: draft-context error checks (#24922), remote preset fix (#24938), per-user-message checkpoints (#24176)
• vulkan / sycl / opencl / hexagon: broad backend fixes and op coverage (FA bias-before-softmax overflow fix #24909, GET_ROWS_BACK #24883, CONV_3D #24612, hexagon MUL_MAT rework #24954, etc.)

Full list in CHANGELOG.md.

Verification

• Full test suite: 158 passed, 4 skipped
• Smoke tests (--include smoke): 7 passed — generation, streaming, chat templates, JSON-schema grammar, raw GBNF, embeddings (Qwen3-Embedding model)
• mix format --check-formatted: clean
• mix dialyzer: 0 errors

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