docs: add staged adoption guide - #415

Implements guide 5 from [#390](#390), which asks
for "a general walkthrough of the notebook, fitting calibration temperatures, and pushing a
checkpoint to the Hub, alongside the existing browser-agent example". Nobody has claimed it in
the thread: @emiliano-go is not working on any of the six, and @Bruce-Yii took the command
line + MCP guide (with #415 for staged rollout).

Every number on the page is from committed material -- the README's 0.36/0.35/0.766 and 4-5 h
runtime, the notebook's own cells (slice sizes, hyper-parameters, the 4-6 minute demo run, the
fixed comparison rows) -- or a public issue number. Nothing is estimated, and no training was
run to write it.

## What the page covers

- **The notebook end to end**, as a table: what each step does, the Kaggle settings, and where
  the outputs land.
- **The training recipe**: gold distributions as targets, the GRPO-style policy gradient with
  proper-scoring rewards, the soft cross-entropy term, and the knob table for a 16 GB card.
- **Calibration is part of the run**, the page's centre: the slice is held out *before* the
  shard (fitting on trained items returns a degenerate scale), the fit is per question type and
  clamped, and the inherited `temperature_by_options` is removed in the same write because it
  would take precedence at inference. `tests/test_calibration_persistence.py` is the regression.
- **Evaluating**: the test-split pass, the metric list, the fixed reference rows, and the two
  habits -- hold out your own slices, report calibration next to accuracy.
- **Pushing to the Hub**, and **what to watch** (targets as the ceiling, the small calibration
  slice, primitive limits, config persistence).

Deliberate choices tested:

- **One nav edit.** The Fine-tuning section exists and already holds the browser-agent page, so
  `finetune.md` is listed beside it rather than left to the Guides glob -- #390's rule is that
  an unlisted page shows up under Guides and can be "placed in another section when needed".
- **No link out of the docs tree, and none into pages still in review.** `BENCHMARKS.md` is
  plain text for the same reason as the Routing guide (zensical `--strict` resolves relative
  links), and links to the Questions and answers guide (#418, in review) failed both gates --
  zensical aborted, and `test_packaging.py` named the missing target -- so the primitives'
  limits are linked as issues (#394, #377) instead.
- **It expands the README's Fine-Tuning section rather than copying it**, per #390 ("the
  README summaries kept in place"): the walkthrough, the calibration order and the publish
  cell are notebook-level detail the README does not carry.

```
$ zensical build --strict --clean        exit 0, no griffe lines
$ python tests/test_packaging.py         all packaging tests passed
$ python tests/test_doc_tables.py        54 passed, 0 failed
$ python -m ruff check laya/ --select=E9,F63,F7,F82,F401,F811 --line-length=120
All checks passed!
$ python -m compileall -q laya/ tests/
```

Known limitation: this is the fine-tuning walkthrough only. The other guides #390 lists --
HTTP API (which should wait for #304/#281 to settle), command line and MCP (@Bruce-Yii; #415
landed for staged rollout), benchmarks -- are not covered here.