fix(docs): clarify scaled dot-product attention rationale

[lucasxing1]

Summary

• clarify why scaled dot-product attention divides logits by sqrt(d_k)
• explain the link between larger unscaled dot products, Softmax saturation, and very small gradients
• keep the Chinese and English Chapter 3 explanations aligned

Rationale

The previous wording stated that scaling prevents gradients from becoming too small, but did not explain the mechanism. The revised text follows the rationale in Attention Is All You Need: as d_k grows, unscaled dot products can grow in magnitude and push Softmax into saturated regions; scaling controls the score magnitude and stabilizes training.

Reference: https://papers.neurips.cc/paper/7181-attention-is-all-you-need.pdf

Validation

• reviewed the Chinese and English changes side by side
• confirmed the branch contains only the two intended documentation files
• ran git diff --check
• no runtime tests were needed for this documentation-only change

Fixes #835