Bump llama.cpp to f708a5b (b9846), release v0.8.31

[nyo16]

Summary

Updates the vendor/llama.cpp submodule from 9bebfcb (b9826) to f708a5b (b9846) — 20 commits — and cuts release v0.8.31.

No NIF changes required. Among the headers the binding compiles against, only common/common.h changed, and its diff does not reach anything the binding consumes:

• a block of internal COM_* logging macros (COM_DBG/COM_TRC/COM_INF/COM_WRN/COM_ERR/COM_CNT);
• a new COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH value appended to the common_speculative_type enum (the DFlash speculative-decoding work);
• common_params_speculative::need_n_rs_seq() extended to reserve a recurrent-state seq for that new DFlash draft type.

The binding only constructs common_params_speculative for its MTP path (types/draft.*) and otherwise calls common_chat_templates_*, common_context_can_seq_rm, common_batch_add, and
json_schema_to_grammar — all unchanged. include/llama.h, ggml/include/ggml.h, ggml/include/ggml-backend.h, common/chat.h, common/json-schema-to-grammar.h, common/sampling.h, and
common/speculative.h are untouched.

Notable upstream changes

• model: DeepSeek V4 (#24162); MiniCPM5 chat parser (#24889)
• spec/dflash: DFlash speculative decoding (#22105) + draft-conversion refactor (#25110)
• CUDA/HIP: cudaMemcpy2DAsync fast path in ggml_cuda_cpy (#25057); hipBLAS dense prefill on gfx900 (#24588)
• vulkan: matmul bk-loop roll for Asahi Linux (#24663); flops-based submission heuristic (#25005)
• opencl: flash-attention improvement (#25069)
• ggml-webgpu: NVFP4 support (#25143)
• sched: revert of #20793 (the split-compute sync change that landed in v0.8.30) (#25138)

Verification

• Rebuilt the NIF from source (LLAMA_BACKEND=auto, Metal)
• Full suite: 158 tests + 4 skipped, 0 failures
• All 7 end-to-end smoke tests pass (generation, streaming, chat templates, JSON-schema grammar, raw GBNF, embeddings) — generation against Qwen3.5-0.8B, embeddings against Qwen3-Embedding-0.6B
• mix format --check-formatted clean
• Dialyzer: 0 errors

checksum.exs is intentionally untouched — CI regenerates it after the release tag.