Cleaning it is a bad idea when building from a release tarball.

Reported-by: @alexanderlaw
Closes #19

filter.sql was replaced by known.txt and known_re.txt.

See https://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=fd1a421fe66173fb9b85d3fe150afde8e812cbe4

Blind attempt at increasing generation speed.

Per input from @mlkersten.

This reduces the MonetDB error rate from 90% to 50%.  MonetDB doesn't
appear to support LIMIT clauses in atomic subselects.

The previous commit surprisingly yielded a CPU hogging sqlsmith with
monetdb due to the table_or_query_name production hardly finding
candidates.  It turned out that the scope model was seriously lacking:
Creating a child scope now automatically includes the table and column
references available to the parent scope.  Previously, child scopes
only contained table and column references explicitly put there by the
parent production.

This special-casing for MonetDB is no longer needed as the impedance
matching code will blacklist the production.

Throw the correct exception classes and fill them with sensible data.
Also avoid writing anything to stdout from the DUT class.  If the
previous, extremely verbose output is desired, a logger class should
be implemented that is made available with a new option, e.g.,
--log-to-file.