fix(vllm_plugin): fix multimodal registration for offline inference

[baotran152]

• Wrap dummy audio data in AudioProcessorItems + MultiModalDataItems in
  get_dummy_processor_inputs; plain list caused AttributeError on
  get_processor_data() during KV cache profiling
• Fix mm_data to mm_data_items due to mismatches with vLLM
• Add get_data_parser() override on VibeVoiceProcessingInfo with
  target_sr=24000; previous method was on wrong class and never called,
  leaving resampler with target_sr=None on every generate() call"