Fix notebook content accuracy

[leo-lilinxiao]

Summary

• clarify KV Cache complexity in the inference acceleration and long-context lessons
• fix two English notebook code issues that prevented cells from parsing/running
• replace several placeholder exercise sections with concrete practice tasks
• update notebook counts and curriculum links to reflect the current 26-notebook layout
• add tiktoken to requirements for the tokenizer basics notebook

Why

A few tutorial sections used shorthand complexity claims that could mislead readers: KV Cache avoids recomputing historical K/V and hidden states, but the cache itself grows linearly with sequence length and decode still reads historical KV. The English notebooks also had a couple of small syntax/content artifacts that made the lessons less reliable.

Validation

• git diff --check
• parsed all notebook code cells with Python ast.parse
• npm run build

[leo-lilinxiao]

@sanbuphy 你好！这个教程写得非常好，结构清晰，也很适合从零理解 LLM 核心组件，非常感谢开源这套内容～

这次 PR 主要修了一些内容准确性和英文版中的小问题，例如 KV Cache 复杂度表述、英文 notebook 里的中英文残留/语法问题，以及少量依赖和目录数量同步。我已经 rebase 到最新 main，并重新跑过 notebook 解析、git diff --check 和 npm run build。

辛苦有空帮忙检视一下代码，看看是否合适合入。谢谢！

[sanbuphy]

LGTM 谢谢贡献