Bump llama.cpp to b10830 (465e49b9c, upstream v0.4.0)

[nyo16]

165 commits past b10665 (ca3d5a3e1). LLAMA_COMMIT in the Makefile moves
with the submodule. No NIF source change.

API surface

• include/llama.h: only llama_tensor_read_lazy → llama_lazy_mode
  (tensor_read_lazy → lazy_mode in llama_model_params, #27969). The
  NIF neither sets nor reads it.
• common/common.h: same rename plus preserve_reasoning_specified.
• ggml-backend.h, ggml-rpc.h, chat.h, json-schema-to-grammar.h,
  speculative.h: unchanged.
• llama_model_default_params() / llama_context_default_params():
  value-identical to b10665 apart from the rename.

Upstream defects we work around — all three still stand (source diff)

• ggml_backend_rpc_start_server still returns void.
• ggml_backend_cuda_comm_init untouched; RPC set/get_tensor_2d hooks
  still NULL. RPC diff is #26500 (cross-server buffer serialisation) and
  #27960 (ggml_op_alloc_size_may_expand).
• ggml-cpu/CMakeLists.txt diff adds iqp.cpp and gates SpacemiT IME
  sources; the -mcpu=native probe is untouched.
• :row split mode still throws on CUDA (no ggml_backend_split_buffer_type).

Behaviour change worth knowing

• TENSOR_READ_LAZY tensors (Gemma 4 per-layer embeddings, Qwen4-exp PLE)
  now map the file regardless of load_mode (#27837). :use_mlock alone
  no longer guarantees "no mapping" for those two tensor kinds; every other
  tensor behaves as Model.load/2 documents.

Verification (macOS, Metal, M1 Max, LLAMA_BACKEND=metal)

Run	Result
mix test (no model)	428 passed, 149 excluded
--include smoke embeddings slow mtp (Qwen3.5-0.8B-UD-Q4_K_XL, Qwen3-Embedding-0.6B-f16, Qwen3.6-35B-A3B-MTP-UD-Q4_K_XL)	569 passed, 8 excluded
--include mtp_sidecar (Qwen3.8-27B-Q4_K_M + mtp-Qwen3.8-27B-Q4_0)	434 passed, 143 excluded
--only mtp_cancel	fails as documented: verify decode failed: code=-1, no abort

mix compile --warnings-as-errors and mix format --check-formatted clean.

Not run: rpc_live (needs an RPC build and a worker); the Hex tarball
source-build check.

Files

• vendor/llama.cpp → 465e49b9cea78a68b9c244ffb48d0ee24a82873d
• Makefile: LLAMA_COMMIT
• CHANGELOG.md: bump bullet, re-verification counts, mtp_cancel state
• docs/release-guide.md: defect re-check at b10830
• lib/llama_cpp_ex/model.ex: split-mode pin reference