Building without extensions has been long broken and nobody really cares
about a pure-DBAPI implementation (which could be created using a wrapper
instead).

Unfortunately PQcancel blocks, so it's not better than PQgetResult.
It has been suggested to use PQreset in non-blocking way but this would give
the Python program the burden of handling a connection done but not configured
in an unexpected place.

Otherwise the connection won't be usable in case we manage
to put it back on track (libpq reports "another command is
already in progress")

If the network is down, trying to read blocking will hang the process hard
(ctrl-c not working). Send a cancel signal instead (as suggested in
http://archives.postgresql.org/pgsql-hackers/2012-07/msg00903.php) and go
back into a green polling: this should allow a further error (e.g. another
ctrl-c) to break the loop. In this case we cannot assume anything about
the state of the connection, so we close it.

Use the async_cursor property to store an indication that something is
running (even if it is not necessarily a cursor running the query).

.c files only need to import psycopg.h: it will in turn import
dependencies from Python and libpq and configure.h. psycopg.h should be
the first to be imported, so the basic imports are not required in
the .h's

As a guideline I'm trying to import from the most specific to the most
generic to detect missing imports in the .h's.