core: make sure we never incref events more than once and never decref them more than we increfed them. Thanks to Cyril Bay.

This fixes Cyril's test as well as test__server_pywsgi.py/test__server.TestDefaultSpawn.test_error_in_spawn which DECREFed event twice (detected by python-dbg).

socket: make recv(), send() and other methods keep the reference to socket so that fd is not closed while we're waiting

- also expect waitxxx() to be cancelled and exit appropriatelly in such case

socket: if socket closed during connect() in another greenlet, make connect() raise socket.error(EBADF)

socket: if socket closed during accept() call in another greenlet, make accept() raise socket.error(EBADF)

socket: make wait_read/write/readwrite functions return True if they return normally and False if they were cancelled with cancel_wait()

add server_implements_pipeline option to test__pywsgi.py which disables tests that depend on pipelining working in test__wsgi.py

- core.http: remove unused '_gencb' property
- core.http: make 'handle' property public
- core.http: add default_response_headers argument and property that is passed to each http_request

the old way of settings _default_response_code is no longer supported by core

turns out os.stat(..).st_mtime is float

to avoid confusion with Greenlet's spawn() classmethod

disable test_notexistent, because it takes way too much on windows for some reason and doesn't test much

prevent bogus complains about libevent version mismatch due to invalid VERSION define in libevent's config.h on windows

- the error was only seen with libevent-2 but not with libevent-1.4