feat(sdk): add ignore_eos to force full-length decode in geniex-bench

[RemiliaForever (RemiliaForever)]

Summary

Adds an --ignore-eos option to geniex-bench that keeps generation running
past EOS/EOG so decode always produces the full --n-gen tokens. This makes
decode-speed numbers comparable across models that would otherwise stop early,
mirroring llama-bench --ignore-eos.

The whole behavioral change is one guard in the llama_cpp decode loop; the rest
is plumbing a single bool through the one public conduit that reaches it.

Changes

• sdk/include/geniex.h — append bool ignore_eos to geniex_GenerationConfig (field appended at the end, so existing offsets are unchanged).
• sdk/plugins/llama_cpp/src/{llm,vlm}.cpp — guard the llama_vocab_is_eog break with !cfg.ignore_eos. When set, the EOG token falls through, the loop runs to max_tokens, and stop_reason becomes "length" via the existing post-loop path. Both the text-LLM and VLM decode loops are covered.
• sdk/benchmark/benchmark.c — --ignore-eos flag → options_t → fill_gen_config, plus help text.
• FFI mirrors (required by CONTRIBUTING.md §4 — public header changed):
  • Python ctypes _types.py — append ('ignore_eos', c_bool) (correctness-critical: the SDK copies *config by value, so a shorter Structure would read past the allocation).
  • Go common.go — IgnoreEos bool + wire into toCPtr().
  • Android JNI jniutils.cpp + Kotlin GenerationConfig.kt — read ignoreEos.

qairt

qairt drives EOS inside the third-party genie engine
(third-party/geniex-qairt/.../llm_model.cpp), which has no ignore-eos seam and
must not be modified. The flag is therefore a documented no-op on qairt — the
plugin simply ignores the new field, consistent with how it already ignores
--ngl / --n-ctx / --stop.

Test plan

• geniex.h field compiles; a C consumer can set g.ignore_eos = true.
• benchmark.c passes gcc -fsyntax-only.
• Python _types.py loads; geniex_GenerationConfig._fields_ ends with ignore_eos and set/get round-trips.
• bindings/go/common.go is gofmt-clean.
• clang-format, ruff check, ruff format --check clean on touched files (pre-existing unrelated format drift in llm.cpp left untouched).
• Full SDK bridge build + on-device bench run (--ignore-eos yields gen == n_gen and stop_reason == "length") — pending, needs the toolchain container / device.