Add speculative decoding docs and tests

[wirybeaver]

Summary

• add the Week 3 Day 4 speculative decoding tutorial and book navigation
• add deterministic tests for speculative accept/reject, rewind, EOS, and cache release behavior
• correct the roadmap interpretation: the reference implementation already existed in tiny_llm_ref.speculative_generate; this PR adds the missing docs and tests

TODO

• Run pdm install -v in a local project environment.
• Run pdm run check-installation.
• Run pdm run test-refsol --week 3 --day 4.
• Run pdm run test --week 3 --day 4.
• Run pdm run main --solution tiny_llm_ref --loader week2 --model qwen3-4b --draft-model qwen3-0.6b with local model weights.

Tests Run

• uv run --python 3.12 --with pytest --with mlx --with mlx-lm --with numpy --with setuptools --with nanobind==2.4.0 python -m pytest -q tests_refsol/test_week_3_day_4.py
• python3 -m py_compile tests_refsol/test_week_3_day_4.py
• mdbook build book (passes; existing environment warns that mdbook-katex is not installed and mdbook-toc was built for an older mdbook)

Note: I currently use uv for the targeted pytest run. Will use pdm for verifycation