"Both the privacy and the authenticity properties of OCB degrade as
 per s^2 / 2^128, where s is the total number of blocks that the
 adversary acquires.... In order to ensure that s^2 / 2^128 remains
 small, a given key should be used to encrypt at most 2^48 blocks (2^55
 bits or 4 petabytes)"

-- http://tools.ietf.org/html/draft-krovetz-ocb-03

We deem it unlikely that a legitimate user will send 4 PB through a Mosh
session.  If it happens, we simply kill the session.  The server and
client use the same key, so we actually need to die after 2^47 blocks.

Closes #77.

struct winsize contains fields other than ws_col and ws_row.  To avoid passing
uninitialized data to TIOCSWINSZ, initialize it first using TIOCGWINSZ.

Found by Valgrind.

(closes #85 github issue)

(closes #86 github issue)

(approved by Ted Krovetz, author of OCB reference implementation)

The default is to detect the presence of the library, as before.

Suggested by Michael Weber (github: xmw).

Closes #88.

(closes #84)

On Android we lack libutil, and we configure with --disable-server.

When cross-compiling for Android, configure (erroneously?) decides we don't
have a GNU-compatible malloc, and tries to

    #define malloc rpl_malloc

We don't define rpl_malloc, and it's not clear it would work with the <cstdlib>
C++ header, anyway.  See

    http://nerdland.net/unstumping-the-internet/malloc-has-not-been-declared/

If we aren't using malloc(0), it should be safe to use malloc without the check.

SSE is x86-specific, but we use this function on all archs.

(closes #78)