Previously we cached only the leaf sub-trees of the IR plan under
the first level of motions. The problem was in temporary tables as
we used random names for them and dropped after usage (and invalidated
corresponding prepared statements as well). Now we use idempotent
names for temporary tables and truncate them after usage. It allows
us to keep prepared statements in the cache. One more thing: as we
use the same temporary table for multiple queries, we need to use
latch locks for the prepared statement cache. One day we'll switch
to more granular locks (mostly for vinyl spaces), but let's use
current approach for the start.

We also removed can_be_cached flag from the code. Only DQL uses VDBE
and needs caching, and all these queries can be cached. The only
exception are the queries with Local and LocalSegment motion policies
but only DML can have them. So, no need for this flag and let's make the
code cleaner.