Show both model sizes in efficiency animation

[weitianxin]

Summary

• expand the inference animation from 4B-only to 4B plus 27–30B releases
• normalize each row to its own baseline while retaining exact measured latency and speed-up
• label hardware and fixed panels so cross-hardware latency is not conflated
• update English and Chinese README captions

Validation

• deterministic GIF regeneration
• source values checked against both efficiency result JSONs
• Python compile and git diff --check
• independent visual/data review passed