Refactor WTRecordStoreCursor timestamp fields

Oplog cursors keep around two timestamps as state. The traditional visibility "no holes" time and SERVER-63308 introduced _readTimestampForOplog.

Given visibility should ultimately be the minimum of the two values (when they both exist), it doesn't seem necessary to have two variables. However, an initial effort to make that reduction failed.

I speculate some of the complexity has to do with oplog visibility being reset on _txnClose, but the _timestampReadSource is not. It's unclear what callers rely on those semantics.