The logic for how to check when to update the table cache for
INNODB_LOCKS with real data was flawed. This could result in both
not updating the cache often enough (when the table is queried
repeatedly with less than 100 milliseconds in-between) resulting
in stale data; as well as updating too often (when multiple
queries against the table start at around the same time).

This caused occasional test failures in innodb_information_schema.

Fix by updating the "last updated" timestamp in the right place,
when the cache is updated, not when it is read.