[Example] Add Qwen2.5-Omni offline inference example

[Sakura-Tears07]

Add a vLLM-Omni based example under example/inference/qwen2.5-omni, aligned with the wan2.2-vae layout for MagiCompiler profiling.

MagiCompiler is toggled through VLLM_OMNI_MAGI_COMPILER inside vLLM-Omni.

PR Category

• New Feature
• Optimization
• Breaking Change
• Bug Fix
• Development / Refactoring
• Documentation
• Chore
• Testing

Description

This PR adds a Qwen2.5-Omni offline inference and profiling example based on vLLM-Omni.

Main changes:

• Add the example under example/inference/qwen2.5-omni.
• Add a Qwen2.5-Omni inference wrapper.
• Add warm-up execution before profiling.
• Add repeated inference runs with NVTX ranges.
• Add optional Nsight Systems profiling.
• Support text and video inputs.
• Support local model checkpoints and custom deployment configurations.
• Update .gitignore for generated audio files and profiling reports.

MagiCompiler is enabled using:

VLLM_OMNI_MAGI_COMPILER=1

Set it to 0 to disable MagiCompiler for comparison.