vllm: Fix Qwen3-Omni interleaved M-RoPE boundary positions

vLLM now correctly handles M-RoPE boundary positions for interleaved Qwen3-Omni models. This fix ensures accurate positional encoding across multi-modal sequences, preventing sequence misalignment during inference. The update enables stable deployment of Qwen3-Omni architectures within the vLLM inference runtime.