Decide: introduce a ParametricRegressionFitter base class over the shared fit skeleton

[derrynknife]

Follow-up to #295/#297/#298 (merged in PR #301). The parametric regression families (PH, AFT, PO, AH) now share their fit pipeline via module-level plumbing — prepare_regression_fit / assemble_regression_model / the two optimiser ladders in _fit_skeleton.py — plus HazardIdentitiesMixin and LogLinearPhi. But there is still no RegressionFitter base class analogous to ParametricFitter on the distribution side: each fitter is an independent class composing those pieces.

The decision: whether to formalise a ParametricRegressionFitter base class owning the common __init__ attribute block (dist, k_dist, bounds, support, param_names, param_map, the *_dist aliases) and a template-method fit() with hooks for the optimiser strategy and reg-model builder.

Benefits

• Kills the last drift-prone duplication: the five __init__ blocks are hand-copies and have already drifted in small ways (which *_dist aliases exist, whether the phi signature is validated). The fit-pipeline drift that produced Parametric PH: fixed={"beta_0": ...} silently fixes alpha (param map not offset by k_dist) #251 started the same way.
• Makes the implicit fitter contract explicit — real types instead of fitter: Any in _fit_skeleton.py, isinstance checks, a docs anchor. A new family (e.g. frailty) becomes correct by construction.
• AFT and PO fit() bodies are now line-for-line identical except the reg-model name; a template method removes another ~60 lines.

Risks

• MRO complexity: these classes already stack HazardIdentitiesMixin + TVC mixins + DataFrameRegressionMixin; the TVC mixins call into self.fit internals and a template fit() must not change what they see. This is the most likely place for a subtle break.
• The irregular cases fight the template: PH's public (name, dist, phi, ...) constructor is API and cannot change; AH's optimiser is a penalty loop with a post-fit finiteness raise; the AL parameter-substitution fitter does not fit the shape and should stay outside the hierarchy. A base class with three hooks and two exceptions may just relocate complexity rather than remove it.
• Serialisation coupling: _SERIALISABLE_KINDS / from_dict rebuild fitters via the public factories; constructor-signature or class-identity changes risk breaking round-trips of previously saved models. Keeping constructors frozen avoids this but constrains the design.
• Diminishing returns: the remaining duplication is declarative attribute assignment that fails loudly, not logic that fails silently — the bug-generating duplication is already gone.

If done: roughly half a day; verify with the existing bit-identity reference harness plus the full TVC test suites (the TVC paths are not covered by the reference capture). Deliberately deferred from v0.17.0 since that release already carries a large amount of refactoring.