FEAT: accept array-valued parameters

[grayson-helmholz]

Closes #571

✨ New features

Estimators now accept array-valued parameters. A parameter value of shape (p,) represents p parameter points, and the estimator returns an array of p estimator values — one per point — computed in a single, backend-parallelized call. Scalar parameter values behave exactly as before and still yield a scalar.

The main use case is propagating fit uncertainties: a set of bootstrapped or sampled parameter points can be evaluated in one vectorized call rather than in a Python loop, and on the JAX backend that call is JIT-compiled like any other estimator evaluation.

New public type alias ParameterType in tensorwaves.interface, which is ParameterValue | Array. It is now used wherever a parameter value is accepted: ParametrizedFunction.__call__, Estimator.__call__, and the estimator implementations. ParameterValue keeps its meaning and now documents itself as the scalar parameter type.

Array parameters are given a trailing axis on entry, so that they broadcast against the event axis of a DataSample, and the estimator reductions consequently reduce over the event axis (axis=-1) rather than over the whole array. Estimator.__call__ is annotated as returning float | FloatArray to reflect the new array case.

New public type aliases Array and FloatArray in tensorwaves.interface. Array is np.ndarray[Any, np.dtype[ScalarT]], generic in its scalar type through the new ScalarT type variable, which defaults to Any so that a bare Array stays dtype-agnostic. FloatArray is Array[np.floating], for arrays that have to be real-valued. DataSample is now dict[str, Array].

⚙️ Enhancements

Estimator.__call__ and both estimator implementations are now @overload-ed, so that a mapping of scalar parameter values is typed as returning float, a mapping of arrays as returning FloatArray, and a mixed mapping as float | FloatArray. Callers no longer have to narrow the return type themselves.

Function signatures that necessarily take a real-valued function are narrowed from Array to FloatArray: ChiSquared, UnbinnedNLL, and IntensityDistributionGenerator. This matches what their docstrings already state, namely that they work on an intensity $f_\mathbf{p}: X^m \rightarrow \mathbb{R}$. Signatures that are genuinely dtype-agnostic keep a bare Array, because a lambdified expression may well be complex-valued: ParametrizedBackendFunction, PositionalArgumentFunction, create_cached_function, and SympyDataTransformer.

🔨 Maintenance

All remaining np.ndarray annotations are replaced with the Array alias, so that the array type is defined in one place and can be narrowed per site. Sphinx is taught to resolve the new aliases (Array, FloatArray, ScalarT) through api_target_substitutions and autodoc_type_aliases.

typing-extensions >=4.4.0 becomes a runtime dependency on Python <3.13, because ScalarT needs PEP 696 type-variable defaults.

The callback logs in ScipyMinimizer now reuse the estimator value that was just computed, instead of re-evaluating the estimator for the log and casting the cached value with float(). The cast on the value that is handed back to SciPy itself stays, since SciPy needs a scalar.

Squash commit messages

* ENH: implement `Array` and `FloatArray` type aliases
* ENH: narrow intensity function types to `FloatArray`
* ENH: use `@overload` for estimator calls
* MAINT: remove redundant estimator value casts

[redeboer]

⚠️ Performance Alert ⚠️

Possible performance regression was detected for benchmark 'TensorWaves benchmark results'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 2.

Benchmark suite	Current: 1a4a844	Previous: f5c1db9	Ratio
benchmarks/ampform.py::TestJPsiToGammaPiPi::test_fit[10000-jax]	0.5013097544496096 iter/sec (stddev: 0)	1.439081184425215 iter/sec (stddev: 0)	2.87

This comment was automatically generated by workflow using github-action-benchmark.