Bump llama.cpp to a894dae93 (b11064, +13 from 5b59b83f4), release v0.8.53

[nyo16]

Bumps vendor/llama.cpp from 5b59b83f4 (b11050+1) to a894dae93 (b11064), 13 commits. Makefile
LLAMA_COMMIT follows the submodule, @version → 0.8.53.

Compatibility

• No diff in any header the NIF includes (llama.h, ggml.h, ggml-backend.h, ggml-rpc.h, chat.h, json-schema-to-grammar.h, speculative.h).
  src/llama-model.cpp / src/llama-context.cpp untouched, so llama_model_default_params / llama_context_default_params are unchanged. NIF source untouched.
• RPC protocol stays at 7.0.0; a v0.8.52 worker and a v0.8.53 client interoperate.
• All three upstream defects tracked in docs/release-guide.md still stand (source diff): ggml-cpu/CMakeLists.txt, ggml-cuda.cu, and ggml-rpc.cpp are absent
  from the range, so the workarounds stay.

Upstream changes worth knowing

• Chat parsing: dedicated Ling 3.0 / Bailing V3 parser (#28682) — those templates pre-open the think block, so a tool call before </think> used to be classified
  entirely as reasoning and clients saw empty content with no tool_calls. Gemma 4 required-tool grammar fixed (#29115). PEG parser handles invalid UTF-8 in model
  output by returning the maximal subpart instead of failing the parse (#29161).
• json_schema_to_grammar accepts an escaped hyphen (\-) in a regex pattern (#29127). LlamaCppEx.Grammar.from_json_schema(%{"type" => "string", "pattern" => "^[a-z\\-]+$"}) returned {:error, _} before; it now yields a grammar.
• Metal: dsv4_hc_pre supports arbitrary hc instead of falling back to CPU (#29169, Kimi-K3); FWHT kernel reads F16 input directly (#29094).
• CUDA: sparse flash attention for Qwen4 (#28770).
• Mamba: time-step projection input made contiguous (#28832).
• Hexagon: TOP_K, GEGLU_QUICK, I32 GET_ROWS (#29113, #29114, #29116).

Verification

macOS (Metal), M1 Max:

Suite	Result
mix test (no model)	430 passed, 157 excluded
--include smoke --include embeddings --include slow --include mtp	573 passed, 14 excluded
--include mtp_sidecar (Qwen3.8-27B-Q4_K_M + mtp-Qwen3.8-27B-Q4_0)	436 passed, 6 skipped (E4B pair not on disk)
Hex tarball source build, LLAMA_BACKEND=cpu	clones a894dae93, compiles clean

Models: Qwen3.5-0.8B-UD-Q4_K_XL, Qwen3-Embedding-0.6B-f16, Qwen3.6-35B-A3B-MTP-UD-Q4_K_XL. rpc_live not re-run.

Throwaway probe against the new NIF for the #29127 fix: "^[a-z\\-]+$" and "^a\\-b$" both return {:ok, _} with GBNF matching upstream's expected output
(([a-z\-]+), ("a\-b")).

After merge

git tag v0.8.53 && git push origin v0.8.53

triggers precompile → checksum → hex publish.