Fix for memory leak in dblink

There is a memory leak in dblink when we cancel a query during
returning tuples. It could leak a PGresult because memory used
by it is not palloc'ed one. I wrote a patch[1] before, but I've
badly used global variables to track the resource.

The attached is a cleaned up patch rewritten to use a tuplestore
(SFRM_Materialize mode) to return tuples suggested at [2]. Since
we don't return from the dblink function in tuplestore mode, we
can surely release the PGresult with a PG_CATCH block even on error.

Also, dblink_record_internal() and dblink_fetch() are rearranged
to share the same code to return tuples for code refactoring.

[1] http://archives.postgresql.org/pgsql-hackers/2009-06/msg01358.php
[2] http://archives.postgresql.org/pgsql-hackers/2009-10/msg00292.php

Regards,
---
Takahiro Itagaki
NTT Open Source Software Center