Since ctr_drbg does not implement getstate/setstate, it would raise an
exception.   Resetting it and using the global CTR_DRBG instance for actual
samples is fine since it is seeded from the OS CSPRNG, and the distribution
is the only thing that should be saved.

This replaces Mersenne Twister with a simple CSPRNG based on CTR-AES-128,
similar to the CTR_DRBG from NIST SP 800-90A.  The way it is used in
ScrambleSuit is to seed itself off of the PRNG seed per the ScrambleSuit
protocol (either derived from the long term secret, or receive from the peer
post handshake), generate the Probability Distribution and then immediately
reseed from os.urandom.

 * ACCEPTABLE_CMDS now only contains CONNECT, since that is the only command
   used and actually implemented.
 * Removed trailing whitespace from the socks5 tests.

Not going to touch #10240 for now since the SOCKS args handler will eventually
be totally rewritten (changed to return a dict containing the separated [k=v]
pairs).

 * Since the compat inet_ntop is used, no longer need to try/catch
 * Make _ByteBuffer.get(self, length) return a string and remove a ton of
   str() calls

Apparently, these routines are missing on Windows.

Changed based on feedback from hellias:
 * handleCmdConnectFailure now traps errors
 * sendReply no longer deals with parsing out the outgoing local address
 * logging uses the obfsproxy logger again
 * Fixed a bug where send_reply was used instead of sendReply
 * The obfsproxy specific SOCKSv5 classes were renamed to OBFSSOCKSv5Outgoing
   and OBFSSOCKSv5Protocol

Bugs I found:
 * SOCKSv5Protocol.dataReceived() should actually work now (never called by
   obfsproxy since it's overridden)
 * _ByteBuffer.get_uint32() was always doing byte order conversion (method never
   used)