V08 of the package DDL: when an execution ended.

Two changes, one version:

• ended_at, a nullable utc_datetime_usec on executions, the type V07 gives retired_at on charts;
• a non-unique index on executions(ended_at).

ended_at is written once, by the first terminal write a row takes while the column is NULL, and never again: StatifierPersistence.Storage.Adapter.update_execution/2 keeps a stored stamp over whatever the record it is given carries. Before this version the only time an executions row carried was updated_at, which any later write moves, so "how long has this execution been finished" had no column that answered it and kept answering it. The index is what a query over that column needs to avoid a sequential scan of the whole executions table.

Every row that exists when this version runs gets a NULL, terminal rows included: the version adds a column and backfills nothing, because the time a row that is already terminal ended is not stored anywhere to copy from. Such a row made its transition into a terminal status before the column existed, so it reads NULL until a later terminal write reaches it - a re-delivered settlement recording a child's answer is one - and stamps it with that write's time.

Both changes are ones every backend has, so nothing here is guarded by adapter, and the version runs inside an ordinary transaction. down/1 drops the index and the column, stamps and all.