Further refine multiseries stacking utilities

[exalate-issue-sync]

As an EvalML user, I noticed that when using datasets with underscores in the name, stacking and unstacking multiseries data occasionally causes the column names to disappear.

In addition, given that the series ID order can be inconsistent within a series, I would like to be able to return the series ID column with an outputted prediction.