arima: implement bootstrap prediction intervals in forecast_arima

[Coro (AlejandroCoronadoN)]

Summary

forecast_arima(..., bootstrap=True) raised NotImplementedError("bootstrap=True"), even
though the bootstrap simulation infrastructure already exists
(simulate_arima(..., error_distribution="bootstrap"), which resamples the in-sample
residuals). This PR wires that infrastructure in so bootstrap=True produces empirical
prediction intervals instead of raising. Addresses #1191 (the ARIMA part).

What it does

When bootstrap=True, instead of the parametric pred ± q·se formula, forecast_arima
now simulates npaths future sample paths by resampling the fitted residuals
(simulate_arima(..., error_distribution="bootstrap")) and takes the empirical quantiles
for each requested level. This is the standard empirical/bootstrap prediction-interval
method (Hyndman & Athanasopoulos, Forecasting: Principles and Practice; the same approach
as R's forecast::forecast(bootstrap=TRUE)), implemented natively with the library's own
simulate_arima/sample_errors — no external code.

Notes:

• The caller's future xreg (before the internal drift column is appended) is passed to
  simulate_arima, which adds the drift term itself, so drift is not double-counted.
• A clear ValueError is raised if the fitted model does not store residuals.
• Output shape and columns (["80%", "95%"], etc.) match the parametric branch.

Verification

model = AutoARIMA(season_length=1).fit(y).model_
fb = forecast_arima(model, h=12, level=[80, 95], bootstrap=True)   # no longer raises

• New test test_forecast_arima_bootstrap_intervals: intervals are finite, bracket the
  mean, and the 80% band sits inside the 95% band.
• All 47 existing tests/test_arima.py tests still pass.
• ruff check/ruff format clean on the changed lines.

Follow-up

The issue also mentions the bootstrap argument being unused in the ETS path
(pegelsfcast_C). Wiring bootstrap through the ETS interval computation (routing the
analytic classes through the existing simulation path) is more involved and left as a
follow-up so this PR stays focused on the ARIMA NotImplementedError.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}