fix: share the EP tutorial centre and implement width regression - #75

Tutorial 5 declared a shared centre but assigned independent priors, and its regression factor returned constant -1. Assign the same centre prior to every profile and replace the placeholder with a Gaussian width-relation density. The tutorial now checks that every dataset/regression factor actually updated, and smoke CI runs it with real sampling.

Closes #73.

Scripts Changed

• scripts/chapter_advanced/tutorial_5_expectation_propagation.py: shared centre; explicit FWHM collection; Gaussian width relation; explanation of the modelling assumption and rejected projections; stale-factor guard.
• config/build/no_run.yaml: remove the obsolete exclusion.
• config/build/profile_smoke.yaml: use PYAUTO_TEST_MODE=0 for this tutorial so constant likelihoods cannot hide behind sampler bypass.
• Regenerated notebook and navigator catalogue through PyAutoHands.

Modelling choice for review

The formerly unexplained regression is now explicitly a teaching assumption: FWHM_i scatters around m*i+c, using dataset index i=0,1,2 and fixed scatter 5 pixels. These are inferred widths, not invented extra observations. The c prior is centred on FWHM(sigma=10), with sigma=10 pixels; m has mean=0 and sigma=5. The author should review whether this is the desired pedagogical example. Five sweeps demonstrate updates, not guaranteed convergence. No upstream library API change.

Validation

• Real headless sampling exited 0, no STALE FACTORS warning.
• Additional bounded witness execution (not an independent review): shared prior identity across all three datasets; normalized Gaussian log density at an exact relation; changing m, c or a width lowers that density.
• Inspected generated model and state figures: blue shared x3 centre; 5 updates / 5 sweeps for dataset_0, dataset_1, dataset_2 and linear_regression; 0 stale. One marginal projection was rejected in the seeded witness and is shown honestly as reverting; no convergence claim.
• Full smoke discovery: 17 passed, 2 failed solely because blackjax was absent. Installed the declared BlackJAX dependency and reran both failures plus final EP script: 3 passed, including final stale-factor guard and real-sampling smoke override. Thus all 19 discovered runnable scripts were exercised successfully across these runs.
• Regenerated notebook includes revised code/prose; catalogue diff is only the added Width Relation section; git diff --check passed.
• Heart vitals: STALE, published reason release validation stale: source moved since rehearsal (PyAutoFit, PyAutoGalaxy, PyAutoLens). No release claimed.

Part of four authorized linked teaching tasks. Stop at human review; no merge authorization. Runtime-guidance PR #74 is independent; this PR starts from main.

CI prerequisite found during validation

Both Python 3.12 and 3.13 smoke jobs on the runtime-only PR #74 fail because BlackJAX is absent for gradient tutorials 6/7. Add the already-declared blackjax>=1.6.2 dependency to the workspace-owned .github/scripts/smoke_install.sh; no workflow or Python-matrix change. The two tutorials pass locally with this dependency installed. This prerequisite belongs here with restoration of smoke coverage; #74 should be revalidated after this PR lands.