docs(t2i): add Qwen-Image example to text-to-image best practice

[Bruce-Yii]

Summary

The text-to-image best practice (t2i_eval.md) only covers FLUX.1-dev and HiDream-I1-Dev. Qwen-Image is the major domestic open-source text-to-image model, but users have no in-repo example for evaluating it with EvalScope.

This PR adds a Qwen-Image example (zh + en) to the best practice doc, following the existing FLUX.1-dev pattern.

Changes

• docs/zh/best_practice/t2i_eval.md: Qwen-Image example (Qwen/Qwen-Image, QwenImagePipeline, bfloat16, EvalMuse)
• docs/en/best_practice/t2i_eval.md: same example in English

Notes on the example parameters:

• num_inference_steps=50 follows the diffusers QwenImagePipeline default.
• QwenImagePipeline ignores guidance_scale; classifier-free guidance uses true_cfg_scale + negative_prompt, which are forwarded to the pipeline through generation_config (GenerateConfig allows extra keys, Text2ImageAPI passes model_extra through).

Verification

• Markdown fences balanced in both files; all Python blocks parse (ast.parse).
• Docs-only change; no code, no benchmark metadata, no auto-generated files touched.

Fixes #980

[Yunnglin]

LGTM — thanks for the clean, well-scoped docs addition (closes #980).

Verified against the diffusers QwenImagePipeline docs and our source:

• num_inference_steps=50 and true_cfg_scale=4.0 match the pipeline defaults;
• guidance_scale is indeed a no-op for this pipeline; CFG correctly uses true_cfg_scale + negative_prompt;
• the true_cfg_scale / negative_prompt passthrough works via generation_config (GenerateConfig extra=allow -> model_extra -> Text2ImageAPI.generate). Because a non-empty generation_config is provided, our default guidance_scale=9.0 is not injected, so the example is safe as written.

Pushed one small follow-up commit (ba5952e) adding the Qwen-Image Technical Report as reference [6] in both zh/en, plus the matching inline [6] marker in the zh intro, so Qwen-Image is cited consistently with FLUX/HiDream. CI is green — merging.