test(eval): add flashinfer regression and benchmark coverage

[lesj0610]

Summary

• refresh the eval/regression branch onto v0.0.24
• update eval/perf.py for the current runtime/backend flow
• add gateway smoke, multilingual quality smoke, Qwen3.5 architecture smoke, and MLA support matrix documentation
• switch compare_q_exllamav3.py and ppl.py to the backend-auto path used by the refreshed runtime

Notes

• This PR is evaluation/documentation only
• No runtime, kernel, or conversion logic is modified here

Validation

• python -m py_compile eval/compare_q_exllamav3.py eval/perf.py eval/ppl.py eval/gateway_regression_smoke.py eval/quality_regression_en_zh.py eval/quality_smoke_multilingual.py eval/smoke_qwen3_5_arch.py

[lesj0610]

Refreshed onto v0.0.24. This branch is now limited to eval/regression/documentation coverage for the current runtime/backend flow.