Make compilation error readable

[gwtnb]

Python 3 prints a line feed in bytes as \n and e.output is bytes, so e.output should be converted to str for readability.

[niboshi]

I think it's unsafe to assume nvcc output as UTF-8 encoded (See #378).

[gwtnb]

Is it safe if 'UTF-8' is replaced with locale.getdefaultlocale()[1] as you wrote at stackoverflow?

[niboshi]

Actually I'm not sure...
Some one reports locale.getdefaultlocale is not robust: https://stackoverflow.com/a/25691701/402996

sys.getdefaultencoding and sys.stdout.encoding are confirmed not working in one of our member's environment.

[unnonouno]

How about adding errors='ignore' ?

[niboshi]

How about errors='replace'? 'ignore' will simply delete foreign characters.
With 'replace' you can at least see there were some characters which has failed to be decoded.

>>> 'abc[あいう]def'.encode('sjis').decode('utf8', errors='ignore')
'abc[]def'
>>> 'abc[あいう]def'.encode('sjis').decode('utf8', errors='replace')
'abc[������]def'

[niboshi]

LGTM!