feat(mtmd): add Gemma4 audio SMT support

[muggle-stack]

Overview

This PR adds Gemma4Audio support to the SMT audio path.

• Adds Gemma4 dynamic-shape ONNX audio encoder routing through audio_model.encoder_model_path.
• Wires Gemma4 audio prompts through the SMT CLI and llama-server audio request path.
• Uses ORT CPU for the Gemma4 dynamic encoder session; the existing Qwen3-ASR split frontend/backend EP path is unchanged.
• Keeps unsupported fixed-shape bucket selection out of the runtime path. The model directory should point encoder_model_path at the dynamic audio_encoder.q.onnx.

Validation

K3 clean build from a fresh synced source tree:

cmake -B build-dynamic-smt -S . \
  -DGGML_CPU_RISCV64_SPACEMIT=ON \
  -DGGML_RV_ZBA=ON \
  -DLLAMA_SERVER_SMT_VISION=ON \
  -DSPACEMIT_ORT_DIR=/root/gemma4-asr/spacemit-ort-native \
  -DCMAKE_BUILD_TYPE=Release \
  -DLLAMA_BUILD_TESTS=OFF
cmake --build build-dynamic-smt --target llama-server -j8

Runtime command:

build-dynamic-smt/bin/llama-server \
  -m /root/gemma4-asr/gemma4-models-final/gemma-4-E2B-it-Q4_0.gguf \
  --media-backend smt \
  --smt-config-dir /root/gemma4-asr/gemma4-models-final \
  --host 127.0.0.1 --port 18080 \
  -t 8 -c 4096 --warmup

Model config used the dynamic encoder only:

"encoder_model_path": "audio_encoder.q.onnx"

Performance

Measured on k3-2 with llama-server warmup enabled. The first audio request includes ORT CPU encoder session creation and encoder warmup; subsequent requests are the steady-state path.

Case	Audio duration	Elapsed	RTF	Result
First short EN ASR request	1.358s	10.481s	7.716	What is the date today?
ZH ASR, steady-state run 1	14.158s	14.745s	1.041	Completed
ZH ASR, steady-state run 2	14.158s	14.674s	1.036	Completed
ES -> EN translation	3.520s	3.711s	1.054	Excuse me, could you tell me how to get to Central Park?

Long ZH timing breakdown from server logs after warmup:

• audio encode: ~2.61-2.67s / 356 audio tokens
• audio prefill: ~5.34s / 356 audio tokens
• decode: ~6.35s / 70 output tokens
• total: ~14.66-14.73s

Notes

SpaceMIT EP was tested with the dynamic Gemma4 encoder and is not used for this path: EP 8-thread long-audio inference triggered a libspacemit_ep.so.2.0.5 SIGSEGV, and EP 1-thread returned std::bad_alloc during the audio request path. Keeping the dynamic Gemma4 encoder on ORT CPU avoids the crash while preserving the SMT server integration and existing Qwen3-ASR EP behavior.