Description

Fixes some logic in the groupby/sort/sort_helper.cu to improve memoization safety. Some functions cache their results so they may be called several times. This change ensures the cached variable is fully formed before returning it so that parallel calls do not get a partially built result. The code also simplifies some of the logic by replacing some thrust calls with equivalent cudf::detail function calls that return columns.

Found while working on #13489

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.