feat: cross-Analysis shared per-evaluation state in FactorGraphModel

[Jammy2211]

Summary

Adds an opt-in mechanism for the per-factor Analysis objects of a FactorGraphModel to compute a model-dependent object once per likelihood evaluation and reuse it across every factor, instead of each factor recomputing identical work. This is the generic, domain-agnostic foundation for the ALMA datacube likelihood (sub-task B, tracked separately), where ~97% of each channel's inversion-setup work is identical across channels and currently rebuilt N times.

The design mirrors the existing EPAnalysisFactor precedent (graph-level state attached around the factor loop) and Analysis.modify_before_fit (the precompute-then-reuse hook), but runs once per likelihood evaluation rather than once per fit. It is fully backward compatible: when no factor opts in, every graph behaves exactly as before.

Resolves part of #1307 (autofit deliverable; the 1D Gaussian workspace tutorial and fast tests follow).

API Changes

All changes are additive and backward compatible — no removals or renames.

• Added Analysis.shared_state_from(instance) -> None: an opt-in hook (default returns None) that computes a per-evaluation object shared across a FactorGraphModel's factors. The per-evaluation, cross-factor sibling of modify_before_fit.
• Analysis.log_likelihood_function (and the AnalysisFactor / PriorFactor / HierarchicalFactor variants) gain an optional shared=None keyword argument. Existing analyses that do not declare it keep working because FactorGraphModel forwards shared only when a factor opts in.
• af.ex.Analysis gains an opt-in share_model_data=False constructor flag plus a worked shared_state_from implementation, demonstrating the mechanism on the 1D Gaussian toy.

See full details below.

Test Plan

• python -m pytest test_autofit/graphical/ — 188 existing + 4 new tests pass
• New test_autofit/graphical/test_shared_state.py:
  • shared state computed once per evaluation (not once per factor)
  • shared-path likelihood exactly equals the unshared per-factor sum
  • a graph with no opt-in factor is unchanged
  • the default shared_state_from returns None

(Note: the full test_autofit/ suite has 13 pre-existing failures under test_autofit/non_linear/search/nest/nss/ that are unrelated to this change and already broken on main.)

Full API Changes (for automation & release notes)

Added

• autofit.non_linear.analysis.analysis.Analysis.shared_state_from(self, instance) — returns None by default; override to share per-evaluation state across FactorGraphModel factors.
• autofit.example.analysis.Analysis.__init__(..., share_model_data=False) — opt-in flag enabling the shared-state path on the 1D Gaussian example.
• autofit.example.analysis.Analysis.shared_state_from(self, instance) — worked example: computes the shared model data once when share_model_data=True.

Changed Signature (backward compatible — new optional kwarg with default)

• Analysis.log_likelihood_function(self, instance) → (self, instance, shared=None)
• AnalysisFactor.log_likelihood_function(self, instance) → (self, instance, shared=None)
• PriorFactor.log_likelihood_function(self, instance) → (self, instance, shared=None) (accepts and ignores shared)
• HierarchicalFactor family log_likelihood_function(self, instance) → (self, instance, shared=None) (accepts and ignores shared)
• autofit.example.analysis.Analysis.log_likelihood_function(self, instance, xp=np) → (self, instance, shared=None, xp=np)

Changed Behaviour

• FactorGraphModel.log_likelihood_function now asks the lead factor for a shared object (_shared_state_from) before the per-factor loop and forwards it to each factor when non-None. When no factor opts in (the default), behaviour is unchanged.

Migration

• None required. Existing Analysis subclasses and graphs work unchanged. To opt in, override shared_state_from to return a shared object and have log_likelihood_function(self, instance, shared=None) consume it (falling back to its own computation when shared is None).

🤖 Generated with Claude Code

[Jammy2211]

Workspace PR: PyAutoLabs/autofit_workspace#69

[Jammy2211]

Workspace PR: PyAutoLabs/autofit_workspace_test#31