Bump llama.cpp to b10582 (e85caa81e)

[nyo16]

Bumps vendor/llama.cpp from b10435 (9e40df63b) to
b10582 (e85caa81e), 147 commits. Makefile's LLAMA_COMMIT moves with it, so a Hex source build clones the same
tree a git checkout uses.

No API change reaches the binding

c_src/llama_cpp_ex/llama_nif.cpp is untouched, and that is a measurement, not an assumption. Every header the NIF includes is byte-identical across the range:

Header	Diff
include/llama.h	empty
ggml/include/ggml-backend.h	empty
common/chat.h	empty
common/json-schema-to-grammar.h	empty
common/speculative.h	empty
ggml/include/ggml-rpc.h	RPC_PROTO_MINOR_VERSION 0 → 1

The RPC bump (5.0.0 → 5.1.0) is use_count propagation so backends can fuse ops behind RPC (af5172627). Same major, no signature change, and nothing in this repo pins the
protocol version.

llama_model_default_params() / llama_context_default_params() are unchanged, so no default shifted under a field we leave alone — the silent failure mode this checklist exists
for.

Two upstream MTP commits, both no-ops for us

• f466cfa38 (spec : avoid binding reference to null pointer) skips a null dp.result when dp.drafting is false. The NIF sets dp.drafting = true on the seq immediately
  before every common_speculative_draft, so it never took that path.
• 2c6b141ef (common : fix draft-mtp with embeddings) forces embedding = false / pooling_type = UNSPECIFIED inside common_base_params_to_speculative. The NIF does not call
  it — it builds target and draft contexts itself.

Upstream defects: all three still stand

Re-checked as a source diff, per docs/release-guide.md:

1. ARM -mcpu=native probe — ggml/src/ggml-cpu/CMakeLists.txt did change, but only OpenMP target variables, KleidiAI SME2 GEMV sources, and IntelLLVM fast-math gating.
   Nothing near the probe, so LLAMA_CPU_ARM_ARCH stays.
2. -sm tensor with a non-CUDA member — ggml_backend_cuda_comm_init untouched. #26502 moved the tensor-split meta backend and was reverted in f20395dae.
3. ggml_backend_rpc_start_server returns void — signature unchanged, so rpc_preflight_bind / rpc_wait_until_listening stay.

The guide's re-check note now records e85caa81e and, for the first time, a defect whose file did move — with why the movement is irrelevant.

Verified

macOS 25.5 / M1 Max, Metal, no RPC, at e85caa81e. Models from ~/Downloads, one run per tag as test/test_helper.exs documents:

Run	Model(s)	Result
mix test	none	428 passed, 149 excluded
--include smoke --include embeddings	Qwen3.5-0.8B-UD-Q4_K_XL, Qwen3-Embedding-0.6B-f16	548 passed
--include mtp	Qwen3.6-35B-A3B-MTP-UD-Q4_K_XL 3;176m	439 passed
--include mtp_sidecar	Qwen3.8-27B-Q4_K_M + mtp-Qwen3.8-27B-Q4_0	434 passed156;163;176m
--include slow	Qwen3.5-0.8B-UD-Q4_K_XL	438 passed

Plus mix format --check-formatted clean, mix dialyzer 0 errors, and the Hex source build path: mix hex.build → fresh extract → LLAMA_BACKEND=cpu mix compile clones and
links e85caa81e, matching the submodule — the check that catches a drifted LLAMA_COMMIT.

Not run: --include rpc_live (needs a reachable worker and an LLAMA_RPC=1 build) and the DGX Spark CUDA legs. The CUDA figure quoted in the changelog is still the b10435
measurement, attributed as such.

Docs fix picked up on the way

docs/release-guide.md step 3 told you to run the suite with LLAMA_MODEL_PATH / LLAMA_EMBEDDING_MODEL_PATH. Those were removed in v0.8.42, so following the guide literally runs
the suite with no model loaded at all and silently exercises none of the inference paths a bump most needs. Replaced with the LLAMA_SMOKE_* vars, one invocation per tag, and a
note on why GGML_METAL_NO_RESIDENCY=1 is there.

Not in this PR

No @version bump and no tag, matching how #86 shipped: the changelog entry lands under ## Unreleased and the release commit does the rest.