Tutorial notebook: generative model → messages → learning → planning

[albertpod]

The current notebook.py is a demo. Add a marimo tutorial that walks through the model top-down:

1. Prior structure (what each factor encodes).
2. One VMP message derived by hand.
3. Full E-step on a toy trajectory.
4. M-step.
5. Planning as inference.

Pedagogical, not a reference.

[albertpod]

Still wanted; the premise is stale — there is no notebook.py in the repo any more.

Partial coverage now exists: docs/index.md covers the generative model, why message passing, and planning as inference, and the README has a How it fits together section walking train → filter → plan. Neither derives a message by hand or shows an E-step on a toy trajectory, which is the pedagogical part this issue is really asking for.

Suggested starting points: jopa/nodes/transition.py for the rules to derive, tests/test_online.py for the smallest working E-step/M-step loop to build the toy example from.

[albertpod]

Folded into #44.