Fixing handling of errors in green communication

Hello,

Ticket #113 shows psycopg blocked because of an error during async
communication with green threads

http://psycopg.lighthouseapp.com/projects/62710/tickets/113-psycopg2-eventlet-connection-hang

We have that clumsy function psyco_clear_result_blocking() in green.c
that tries to put the connection back in a working state after an
error in the callback (whick could be either a python error or more
likely a network problem). But because it is blocking it stops other
greenlets to run. I've tried sending a PQcancel instead but it is
blocking as well. Another suggestion I've been given is to use PQreset
in the nonblocking version, but it would put the program in an
unexpected state: the user would find an established connection but
with eventual configuration made in Python discarded.

So for the moment I think the best solution is to close the connection
if there is an error in the callback (meaning an error in the
communication, not an error in a query: these ones are handled
normally). This would put a Python program in the relatively normal
state of dealing a broken connection: a long-living program may
already be able to cope with this eventuality; a short-lived one would
be at least responsive to ctrl-c (if the network is down the blocking
PQfunctions make the interpreter unresponsive to SIGINT).

Suggestions for better solutions are well accepted. However discarding
a connection when we have lost communication and we don't know in
which state it is anymore seems a reasonable one. The solution
(together with a test program to test with failures in green querying)
is currently living in the fix-113 branch of my repos.

-- Daniele