Closes #3894.

There were two query cache related races with identical manifestation.

1. A simple one. Table and database data modification id was updated before actual data modification, so it was possible to associate an old data with the new data modification id in a query cache.

2. A complex one:

• Connection 1 starts a query and takes snapshot of all related tables.
• Connection 2 modifies a row in this table and commits this modification.
• The query from connection 1 takes data modification ID already modified by connection 2.
• The query is executed over this already outdated snapshot and results are put into cache with new data modification ID and old data.
• Connection 1 executes the same queries again and again, new snapshots contain changes from connection 2, but queries take the same data modification id every time and therefore return an outdated cached result until some other modification arrives.