Bump llama.cpp to ddd4ec142 (b10217), release v0.8.41

[nyo16]

Maintenance bump from b10178 (992c32532) to b10217 (ddd4ec142), 39 commits.

This range does not break the upstream C API — no NIF change was required
and the public Elixir API is unchanged.

API compatibility

Two binding-relevant headers moved, both additive:

• include/llama.h — llama_model_params gains a trailing bool load_mtp,
  so MTP tensors are only loaded when actually used (llama : load MTP tensors only if they are really used ggml-org/llama.cpp#26296).
  The NIF builds its params from llama_model_default_params()
  (c_src/llama_cpp_ex/llama_nif.cpp:329) rather than aggregate-initialising the
  struct, so the new field picks up its upstream default — source-compatible, no
  change needed. A new llama_vocab_get_suppress_tokens accessor
  (llama: move suppress_tokens handling to common/sampling ggml-org/llama.cpp#26276) is not read by the binding.
• common/common.h — the inline common_params_sampling::has_logit_bias()
  helper was removed alongside the suppress-tokens refactor
  (llama: move suppress_tokens handling to common/sampling ggml-org/llama.cpp#26276), and a test-only common_get_model_or_exit was added
  (tests : avoid building get-model.cpp many times ggml-org/llama.cpp#26317). The binding calls neither and does not use
  common_params_sampling, so both are inert here.

ggml/include/ggml.h, ggml/include/ggml-backend.h, ggml/include/ggml-metal.h,
common/chat.h, common/sampling.h, common/speculative.h and
common/json-schema-to-grammar.h are all verified untouched in this range.

Changes

• vendor/llama.cpp — 992c32532 → ddd4ec142 (tag b10217)
• Makefile — LLAMA_COMMIT bumped in lockstep, so Hex source builds clone
  the same revision as a git checkout gets
• mix.exs — @version → 0.8.41
• CHANGELOG.md — new ## v0.8.41 section

checksum.exs is intentionally untouched: the precompile workflow's checksum
job generates it and commits it to master after the tag push.

Verification

NIF rebuilt from source at the new pin (LLAMA_BACKEND=metal), then:

GGML_METAL_NO_RESIDENCY=1 LLAMA_BACKEND=metal \
LLAMA_MODEL_PATH=~/Downloads/Llama-3.2-1B-Instruct-Q4_K_M.gguf \
LLAMA_EMBEDDING_MODEL_PATH=~/Downloads/Qwen3-Embedding-0.6B-f16.gguf \
LLAMA_SMOKE_GEN_MODEL=~/Downloads/Llama-3.2-1B-Instruct-Q4_K_M.gguf \
LLAMA_SMOKE_EMB_MODEL=~/Downloads/Qwen3-Embedding-0.6B-f16.gguf \
  mix test --include smoke --include embeddings

489 passed, 0 failures, 21 excluded — matching the b10178 baseline exactly.

Also clean: mix compile --warnings-as-errors, mix format --check-formatted,
mix credo --strict (no issues), mix hex.build.