Sync all EN notebooks to strict 1:1 mirror of ZH

Rebuilt 22 EN notebooks so every ZH notebook now has a cell-for-cell
English counterpart (same cell count, order, cell_type, and ✅ emoji count):

- part1-foundation: 02-bpe-tokenizer, 03-embedding, 04-position-encoding,
  05-transformer-block, 06-mini-gpt, 07-bert-encoder
- part2-training: 08-gpt2-to-modern-models, 09-model-config, 10-moe,
  12-scaling-laws, 14-data-engineering, 15-lora, 17-rlhf-alignment
- part3-inference: 19-generation, 20-inference-acceleration,
  22-speculative-decoding
- part4-frontiers: 24-long-context, 25-cot-thinking
- part5-production: 28-evaluation, 29-distillation, 30-opd, 31-llm-deployment

The EN files had drifted behind ZH (missing sections, missing exercises,
different checkmark conventions). Each was rebuilt as a 1:1 mirror:
markdown fully translated, code logic preserved verbatim with only
comments/print/docstrings translated, LaTeX and demo data untouched.

Verified: 41 ZH ↔ 41 EN, all cell counts match, all ✅ counts match.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: sanbuphy