Investigate core objectives for time series classification

With the merge of #3616, we allow cv fold test splits as small as a single data point for time series problems (if the forecast horizon is set to 1). This is incompatible with some of our classification objectives, namely AUC, as they require instances of all classes in the test set. We should remove AUC and any other similar objectives from the `core_objectives` listed for time series classification problems.

The `forecast_horizon` parameter was bumped up in a few tests in #3616 to avoid this, this issue should revert those changes as well as fixing the underlying problem.