[BUG] fix forecaster default predict_quantiles for multivariate data

[fkiraly]

This fixes a bug reported by @lbventura in #3105: the default implementation of _predict_quantiles, when _predict_interval is implemented, would return quantiles only for the first variable - which is fine in univariate forecasting, but produces too few columns in multivariate forecasting.

This has now been fixed, the correct columns are now returned in both cases, under the condition that _predict_quantiles is implemented but _predict_interval is not.

Odd that this was not tested so far, I still need to investigate why.

To the bets of my knowledge, we should be testing the proba predict methods with multivariate scenarios?

[fkiraly]

Reason why tests did not catch this:

vectorization takes priority over the default fill in, so forecasters that are not genuinely multivariate will call the _predict_interval from _predict_quantiles via vectorization, which is fine.

The problem with the logic therefore occurs only for genuinely multivariate, probabilistic forecasters, of which there were, so far, none in the code base.

Somewhat surprising since VAR is in the code base, but that does not have probabilistic prediction implemented.

[fkiraly]

Included as part of #3105 and #2925