Improve BabyTorch correctness, book structure, and print editions

[amjadmajid]

What changed

• fixes repeated-backward accumulation so leaf gradients accumulate without reusing stale intermediate gradients
• corrects Adam/AdamW bias correction for parameters that start receiving gradients later
• hardens optimizer, scheduler, DataLoader, and BPE validation, including symbol-safe BPE merging and strict unknown-token handling
• adds regression tests for the corrected behavior
• gives every English chapter explicit learning goals and key takeaways
• adds a preface, PyTorch migration appendix, glossary, and primary-source reading list
• corrects overstatements around PyTorch parity, production GPT tokenization, PPO, and RLHF in the book and tutorials
• improves the corresponding Arabic explanations
• adds reproducible 6×9-inch print interiors with front matter, contents, part dividers, embedded fonts, figure conversion, running heads, and RTL-aware Arabic layout
• reduces the Arabic body typography to 9.5pt with tighter leading following review feedback

Why

The repository already had a strong educational sequence, but several correctness edge cases and production-layout weaknesses undermined its promise that the implementation, tests, and book agree. The revision makes the code safer, the explanations more precise, and both language editions suitable for print.

Validation

• 202 tests collected: 200 passed, 2 hardware-dependent GPU tests skipped
• snippet drift guard passes for English and Arabic
• no broken local links in the manuscript
• English PDF: 124 pages, 6×9-inch trim
• Arabic PDF: 115 pages, 6×9-inch trim
• fonts embedded and representative/all-page visual sweeps completed
• git diff --check, Python compilation, and shell syntax checks pass

Generated PDFs remain ignored build artifacts; the reproducible sources and build pipeline are included.