Bump llama.cpp to 4f37f51 (b9967) for v0.8.35

[nyo16]

Summary

Maintenance release: bumps the vendor/llama.cpp submodule from a646006f0 (b9932) to 4f37f5197 (tag b9967) — 35 upstream commits — and prepares the v0.8.35 release.

API compatibility

No NIF changes required. The only header change in the whole range is additive — a new ggml/include/ggml-et.h for the initial ET backend. Every header the NIF uses is untouched: llama.h, common/chat.h, common/common.h, common/sampling.h, common/speculative.h, common/json-schema-to-grammar.h.

Notable upstream changes

• llama core: all KQ masks f16 when flash attention is used, remove zero attention bias and raw_k repeats in DeepSeek V4 (#25370); static tensor-split regex patterns (#24710); llama-batch unit test (#25471)
• ggml: initial ET backend (#24179); smaller-chunk CUDA ggml_top_k()/ggml_argsort() to reduce temp buffer memory (#24776); fix depthwise conv2d (#25490)
• Metal: CONV_2D_DW (depthwise convolution) support (#21565)
• CUDA: mmvq indexing simplification (#25445); snake fusion matcher alignment (#25460)
• HIP/WebGPU/OpenCL: -funsafe-math-optimizations on HIP (#24668); WebGPU flash_attn_vec subgroup tuning (#25418); OpenCL cluster-parallel decode FA for Adreno (#25473), Q6_K GEMM/GEMV fix (#25464)
• hexagon: ARGSORT perf (#25512); unary op tiling/tracing (#25474); VISION RoPE (#25216)
• mtmd: deepseek-ocr v1 multi-tile (#24717)
• tools/server/ui (not linked into the binding): null sampling params (#25538), min-step prompt batch splitting (#25420), chat-template thinking probe hardening (#24093), and assorted CLI/web UI fixes

Verification

• Clean rebuild from source (LLAMA_BACKEND=metal, --warnings-as-errors) — no warnings
• Full test suite with all five model env vars, smoke tests included: 252 tests, 0 failures
• mix format --check-formatted clean