stdout capturing breaks sys.stdout.encoding

py.test by default replaces sys.stdout with something that captures test output, which is great! But the replacement stdout doesn't have a correct encoding attribute, and that can break the code under test.

(Prior to Python 3.6 sys.stdout.encoding is the only way to figure out the current OEM encoding on Windows, which you need if you have to decode the output of subprocesses. You can see what happens when people try to use py.test on such code here: https://github.com/mgedmin/check-manifest/pull/92)

Here's a reproducible test case for Linux (note that you need to use Python 3.5 or older):

# test.py
import sys

def test_stdout_encoding():
    assert sys.stdout.encoding == 'ANSI_X3.4-1968'

LC_ALL=C py.test test.py -s     # succeeds

LC_ALL=C py.test test.py        # fails, sys.stdout.encoding is 'UTF8'

$ pip list|grep pytest
pytest                        3.9.3        
pytest-catchlog               1.2.2        /home/mg/src/pytest-catchlog
pytest-forked                 0.2          
pytest-xdist                  1.24.0       

This is similar to #2375, except I'm not using the capsys fixture.

Start with the test.py reproduction under Python 3.5 or older, comparing LC_ALL=C with and without -s. Trace pytest's default stdout capture and verify that the captured stream preserves the encoding expected by code under test; done means the reproduction passes without -s.