Use subprocess instead of os.system

[jwodder]

closes #1026, closes #1470, closes #1476

[davidism]

Master is now Python 3 only. This needs to be rebased, and compat for shutil.which can be removed.

#1543 reverts the previous shlex.quote pr because of an issue with calling quote on a command that might actually be a command and options, like less -FRSX (and also an issue with quoting on Windows, which is sidestepped here). Unfortunately, just calling shlex.split on the command first will fail on Windows unless slashes in paths are double escaped:

>>> shlex.split("C:\\path\\to\\exe /test other\\path")
['C:pathtoexe', '/test', 'otherpath']
>>> shlex.split("C:\\path\\to\\exe /test other\\path".replace("\\", "\\\\"))
['C:path\\to\\exe', '/test', 'other\\path']
>>> shlex.split("C:\\path\\to\\exe /test other\\path", posix=False)
['C:path\\to\\exe', '/test', 'other\\path']

Technically that last one with posix=False is not correct, as the argument refers to non-Posix shells on Unix, not completely different platforms like cmd on Windows. Either solution would probably be fine for the limited case here, although I'd lean towards replace since it's more clear what's being special-cased.

[AndreasBackx]

I noticed _pipepager exists now. Should we not use that instead here and remove the pipe argument I introduced?

[davidism]

If that works as well, I'm usually in favor of avoiding new arguments.

[AndreasBackx]

I'll try to test this on Windows and get this merged then.

[AndreasBackx]

It seems that when more is piped into, it adds \r\n on Windows. It does not do this when reading from a file using more < .... I tried some workarounds but failed so it's not using piping with more for that reason.

Additionally, which usage is required as recommended by the Popen Python docs. If not used on Windows, it will results in an OSError because of there being a difference on how the resolving works. I've made the pager methods return True if which was able to find the file and False otherwise. This makes it easy to try other pagers we support easily.

It seems to be that newline behaviour is different on GitHub's test runner however. On GitHub \n is not translated into \r\n but it is on my machine. 🤔