Test self.theclass, not datetime. Regression introduced by manual tests.

of datetime.datetime: microseconds are now rounded to nearest with ties going
to nearest even integer (ROUND_HALF_EVEN), instead of being rounding towards
zero (ROUND_DOWN). It's important that these methods use the same rounding
mode than datetime.timedelta to keep the property:

   (datetime(1970,1,1) + timedelta(seconds=t)) == datetime.utcfromtimestamp(t)

It also the rounding mode used by round(float) for example.

Add more unit tests on the rounding mode in test_datetime.

_PyTime_Divide() rounding was wrong: copy code from Python default which has
now much better unit tests.

Oops, I forgot to document my change.

PyObject_Length() returns a P_ssize_t, not an int. Use a Py_ssize_t to avoid
overflow.

On Windows, the tv_sec field of the timeval structure has the type C long,
whereas it has the type C time_t on all other platforms. A C long has a size of
32 bits (signed inter, 1 bit for the sign, 31 bits for the value) which is not
enough to store an Epoch timestamp after the year 2038.

Add the _PyTime_AsTimevalTime_t() function written for datetime.datetime.now():
convert a _PyTime_t timestamp to a (secs, us) tuple where secs type is time_t.
It allows to support dates after the year 2038 on Windows.

Enhance also _PyTime_AsTimeval_impl() to detect overflow on the number of
seconds when rounding the number of microseconds.

On Windows, the tv_sec field of the timeval structure has the type C long,
whereas it has the type C time_t on all other platforms. A C long has a size of
32 bits (signed inter, 1 bit for the sign, 31 bits for the value) which is not
enough to store an Epoch timestamp after the year 2038.

Add the _PyTime_AsTimevalTime_t() function written for datetime.datetime.now():
convert a _PyTime_t timestamp to a (secs, us) tuple where secs type is time_t.
It allows to support dates after the year 2038 on Windows.

Enhance also _PyTime_AsTimeval_impl() to detect overflow on the number of
seconds when rounding the number of microseconds.

* test_eintr: support verbose mode, don't redirect eintr_tester output into
  a pipe
* eintr_tester: replace os.fork() with subprocess to have a cleaner child
  process (ex: don't inherit setitimer())
* eintr_tester: kill the process if the unit test fails
* test_open/test_os_open(): write support.PIPE_MAX_SIZE bytes instead of
  support.PIPE_MAX_SIZE*3 bytes

Skip test_open() and test_os_open(): both tests uses a FIFO and signals, but
there is a bug in the FreeBSD kernel which blocks the test. Skip the tests
until the bug is fixed in FreeBSD kernel.

Remove also debug traces from test_eintr:

* stop using faulthandler to have a timeout
* remove print()

Write also open and close on two lines in test_open() and test_os_open()
tests. If these tests block again, we can know if the test is stuck at open or
close.

test_eintr: don't always run the test in debug mode.

import_init() imports the "_imp" module, not the "imp" module.

format_stack(), and extract_stack() called without arguments.

format_stack(), and extract_stack() called without arguments.

format_stack(), and extract_stack() called without arguments.

Fix test_write(): copy support.PIPE_MAX_SIZE bytes, not support.PIPE_MAX_SIZE*3
bytes.

sys.stderr is sometimes a StringIO. The redirection was just a hack to see
eintr_tester.py output in red in the buildbot output.

ProactorEventLoop now supports SSL.

Some test_eintr hangs on waiting for the child process completion if an error
occurred on the parent. Kill the child process on error (in the parent)
to avoid the hang.

Reported by Jakub Wilk.

Reported by Jakub Wilk.