Add <forward_subnets> to divert rule in OpenBSD

Fixes bug where all traffic routed to loopback would end up being diverted to the same port.

A possible implementation for the change requested in #94, so that seed
hosts can be used without auto hosts. In this scenario only the
specified hosts (or ips) will be looked up (or rev looked up).

Otherwise the output can vary and confuse our attempts to parse it.

Fixes: 93

We shouldn't come up with a fatal error because of a ENETUNREACH when
trying to contact the DNS server. Although this error shouldn't happen
either.

Fixes #89.

Sometime ago I was in python mode and incorrectly indented a line of the
shell script with spaces instead of tabs. Shame on me. This should bring
things back to their natural order.

Previously the sshuttle shell script would pass the python to use as the
first argument of the command. The new run script no longer does this.
Instead we can obtain the python being used via sys.executable.
Fixes #88.

readthedocs alters docs/conf.py which in turn means python_scm detects a
version and incorrectly adjusts the version number. Here we try to work
around this problem.

We do this by renaming the docs/conf.py file and copying it back to
docs/conf.py when setup.py is invoked. This way, hopefully, scm won't
see the changes to docs/conf.py

References:
http://stackoverflow.com/questions/35811267/readthedocs-and-setuptools-scm-version-wrong/36386177
https://github.com/pypa/setuptools_scm/issues/84

This reverts commit 7ea2d973.

In preparation for new release.

It is often the case that the user has no administrative control over
the server that is being used. As such it is important to support as
many versions as possible, at least on the remote server end. These
fixes will allow sshuttle to be used with servers that have only
python 2.4 or python 2.6 installed while hopefully not breaking the
compatibility with 2.7 and 3.5.

When passing multiple subnet files, e.g., by using -s/--subnets
multiple times or by using it together with subnets passed as positional
arguments append the content from all sources instead of only using the
subnets from the last source. This makes the behaviour of -s/--subnets
consistent with -x/--exclude.

As @brianmay observed in #82 this option is no longer used and can be
dropped.

Fixes the kind of problems reported on #75 but does break the command
line "API" (hopefully).

This allows disabling all client tests using a conftest.py file, if for
example #56 gets merged and the server supports more python versions
then the server.

The server side tests are very incomplete.

Due to nested shells, we need to have multiple layers of quoting. Yuck.

Closes #80

In situations where 2.7 is available and some unsupported 3.x is the
system's default we should probably fallback to 2.7 instead of the
default (that might be e.g. 3.4). This might fix #78.

The fish shell doesn’t support ‘||’ and requires a ‘—python python’
workaround.  This change explicitly calls /bin/sh for the remote shell
commands.

Just because we may have found IPv6 DNS servers from /etc/resolv.conf
doesn't mean we should force IPv6 support.

Instead we should disable the IPv6 DNS servers if IPv6 is disabled.

Note: this will also result in any IPv6 servers specified on the command
line being silently ignored too.

Specifying an IPv6 subnet will still require IPv6 support.

Closes #74

Closes #73.

In some cases (see #43) it seems that some network configurations may
end up setting a skip on lo. As sshuttle adds rules that rely on
filtering/translating packets on lo, this causes problem. This fix
overrides the skip and makes the rules be applied again.
Should fix at least some of the problems reported on #43.

This will make it easier to support other platforms/versions in the
future, e.g., OpenBSD.