Simplify recurrent module (DEVELOPMENT.md §4.C)

[derrynknife]

Address the four simplification points for the recurrent package:

• Extract the copy-pasted multi-start MLE scaffolding from the four repair
  fitters (GeneralizedRenewal, GeneralizedOneRenewal, ARA, ARI) into a shared
  RenewalFitMixin: _multistart owns the start loop, best-start selection and
  the two convergence-failure raises; _bounds_transform owns bounds_convert;
  _attach_inference owns _neg_ll/_mle/_n_obs storage; _initial_dist_params
  de-duplicates the first-event initialisation. Each fitter keeps its own
  likelihood/optimiser call so the numerics are unchanged.

• Make ProportionalIntensityModel inherit RecurrenceSimulationMixin instead of
  the stale pre-mixin simulation copy, gaining seeding, the max_events
  backstop and the data-returning simulators; the covariate vector Z is
  threaded through to the sampler by the public entry points.

• Give the parametric intensity models (HPP/CrowAMSAA/Duane/CoxLewis) and the
  proportional-intensity regression fitters AIC/BIC/standard errors by setting
  _neg_ll/_mle/_n_obs and inheriting LikelihoodInferenceMixin, which is
  generalised off its renewal-only parameter-naming assumption via a
  _parameter_names hook. MSE / from_params models carry no likelihood and
  still raise.

• Add an IntensityModel ABC carrying the iif/log_iif/cif/inv_cif/
  parameter_initialiser contract once, and strip the duplicated docstring
  boilerplate from Duane/CrowAMSAA/CoxLewis (the maths is left distinct).

Add tests for the new inference, regression simulation and ABC contract.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.