Improve test coverage for timestamp races

In WT-4057, we had another report of a race between setting a timestamp and getting a valid transaction snapshot. It would be good to have better test coverage of this combination:

• a multi-threaded workload with checking of on-disk state;
• reading as-of timestamp with separate begin_transaction and timestamp_transaction("read_timestamp=X") calls; and
• exercising of "round_to_oldest" – i.e., a test application where the oldest timestamp can move forward concurrent with a call to timestamp_transaction.