Handle each database in its own transaction in clean

[kakudo415]

Problem

clean lists the target databases, drops them and deletes their metadata rows in a single read-write transaction. Dropping a database is an admin API call, so it is not rolled back when the transaction is aborted. An abort therefore leaves metadata rows for databases that no longer exist. Those rows are listed again on the next run, which tries to drop them again, and once enough rows have accumulated clean never finishes.

Changes

• clean lists the targets in a read-only transaction and then handles each database in its own read-write transaction: re-read the row, drop the database, delete the row. A failure or abort affects only that database, and the rows of databases handled earlier are already committed.
• Inside each transaction the row is re-read and skipped if it was removed or updated after it was listed, so a database that was taken out of the pool in the meantime is not dropped.
• dropDatabase no longer creates an admin client on every call. It takes the admin client from the caller: Pool passes the one it already holds, and CleanAll creates one for the whole run and closes it afterwards.

The first non-NotFound error from DropDatabase still stops the run, and a NotFound is still ignored so that the row of an already dropped database is removed.