[MRG] Copy pipeline confidence intervals before inverse transformation - #627

Pipeline._post_predict writes inverse-transformed bounds into the confidence-interval array returned by its estimator. With pandas 3.1/statsmodels 0.15, that array can be read-only, so predicting intervals through an endogenous transformer raises ValueError: assignment destination is read-only. The unmodified base reproduces the existing pipeline6 CI failure with these dependencies. Writable estimator arrays are also modified in place.

Take one local copy of the interval array before inverse transformation. Four tests cover predict and predict_in_sample with read-only and writable arrays, check exact Box-Cox inverse values, and verify the estimator outputs remain unchanged.

Validation: the original pipeline6 fails on the base; all four added regressions fail before the fix; the complete existing pipeline module plus new regressions pass after it (71 passed). flake8 on both changed files and git diff --check pass. Ten existing convergence/weekday warnings remain.

Tested Python 3.12.14, NumPy 2.5.4, pandas 3.1.0, SciPy 1.18.1, statsmodels 0.15.0 and sklearn 1.9.1. Python sources are from base 4c2dfcc; compiled extensions were reused from the 2.1.1 Windows wheel. The entire ARIMA suite and Linux wheel build were not rerun.

The shared pipeline6 CI failure is also visible in #625 and #626.