feat: add SVG2 attention backend

[yiliu30]

Summary

• register a new SVG2_ATTN diffusion attention backend in the plugin
• add the out-of-tree SVG2 implementation with SDPA fallback, warmup gating, and FlashInfer buffer shims
• add environment defaults and unit coverage for backend config and warmup behavior

Validation

• python -m unittest discover -s tests -p 'test_envs.py'
• python -m py_compile src/vllm_qdq_plugin/svg2_attn/__init__.py src/vllm_qdq_plugin/svg2_attn/backend.py src/vllm_qdq_plugin/svg2_attn/impl.py tests/test_svg2_attn.py tests/test_envs.py

Notes

• tests/test_svg2_attn.py could not be executed in this local environment because importing vllm_omni pulled in torchcodec, which failed to load due missing FFmpeg shared libraries.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.