Bump llama.cpp to b10178 and honour :use_mlock / :use_mmap independently

[nyo16]

Updates vendor/llama.cpp from ff067f76d (b10133) to 992c32532 (b10178) —
45 commits — and moves the Makefile's LLAMA_COMMIT with it so Hex source
builds clone the same tree.

Pinned to tag b10178 rather than the branch tip: the 3 commits past it are
CUDA/SYCL/BoringSSL only, and this project pins to tags.

Header impact

Three binding-relevant headers changed in this range. Two are inert:

• common/chat.h — additive only: a COMMON_CHAT_FORMAT_PEG_MINIMAX_M3
  format for the new MiniMax-M3 parser (#26210).
• common/common.h — common_context_seq_rm / _seq_add / _seq_cp
  became static and moved behind a new common_memory struct (#26221).
  The binding never called them: it drives llama_memory_* directly and only
  reads common_context_can_seq_rm, whose signature and
  common_context_seq_rm_type enum are both unchanged. A new
  COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK value (#25173) is likewise inert —
  the binding sets types explicitly for its MTP path.

ggml.h, ggml-backend.h, json-schema-to-grammar.h, sampling.h and
speculative.h are untouched.

The one that isn't inert

include/llama.h gained LLAMA_LOAD_MODE_MMAP_MLOCK and renumbered
_DIRECT_IO from 3 to 4. LLAMA_LOAD_MODE_MLOCK previously meant "mmap and
mlock" and now means mlock without mmap, with the new _MMAP_MLOCK
covering the combination (#26135).

The NIF selects the mode by name, so the renumbering is transparent and the
redefinition is silent — a plain rebuild would have kept compiling while
quietly dropping the memory map for use_mlock: true.

model_load/10 now maps the two options independently:

options	mode
use_direct_io: true	dio (wins outright)
use_mlock: true, use_mmap: true	mmap_mlock
use_mlock: true, use_mmap: false	mlock (anonymous memory, no mapping)
use_mmap: true	mmap
all false	none

This also retires the documented wart where use_mlock: true, use_mmap: false
memory-mapped the file anyway. The default path is unchanged because
:use_mmap defaults to true; only the explicit
use_mlock: true, use_mmap: false combination behaves differently. Documented
as Breaking in the changelog, with Model.load/2 and Server.start_link/1
docs updated to match.

Changelog entries are folded into the unreleased v0.8.40 section;

• mix compile --warnings-as-errors — clean
• mix format --check-formatted — clean
• mix dialyzer — 0 errors
• Hex source build (mix hex.build → extract → LLAMA_BACKEND=cpu mix compile)
  clones exactly 992c32532, confirming LLAMA_COMMIT has not drifted from the
  submodule
• Full suite with real GGUF models (smoke + embeddings): 486/489 passing

Pre-existing failures — not introduced here

The suite is not green. Four failures were A/B-tested by rebuilding at the old
b10133 pin and reproduce there identically, so they predate this bump:

• ServerSmokeTest "a cached prefix is never reused across cache scopes" —
  same_scope is 0
• ServerSmokeTest "PromptCache.restore/4 … trims the unusable tail when only a
  prefix is reusable" — {:error, :seq_rm_refused}, expected {:ok, 3}
• LlamaCppExTest "server cache_prompt sequential requests don't crash on hybrid
  models" — reused is 0
• :mtp tests hard-abort the VM (not the cosmetic Metal teardown assert):
  find_slot: non-consecutive token position 6 after 4 for sequence 0, then
  ggml-backend.cpp:334: GGML_ASSERT(offset + size <= ggml_nbytes(tensor))
  inside llm_graph_input_embd::set_input under generate_mtp_tokens