Also reduced hash function to 256 bits because our target security level
is now 128 bits.

1. The things we're really trying to protect - contact identities,
message contents, etc - can't be erased from memory because they're
encapsulated inside objects we don't control.

2. Long-term secrets can't be protected by erasing them from memory
because they're stored in the database and the database key has to be
held in memory whenever the app's running.

3. If the runtime uses a compacting garbage collector then we have no
way to ensure an object is erased from memory.

4. Trying to erase secrets from memory makes the code more complex.

Conclusion: Let's not try to protect secrets from an attacker who can
read arbitrary memory locations.

DuplexOutgoingSession flushes its output stream if it's idle for a
transport-defined interval, causing an empty frame to be sent. The TCP
and Tor plugins use a socket timeout equal to twice the idle interval to
detect dead connections.

See bugs #27, #46 and #60.

See bug #27.

The TransportReader/Writer's dispose() method should handle that, and
ConnectionManager is responsible for calling it.

This makes it more likely that connections will be closed cleanly.
However, the interrupt() method is currently ineffective for incoming
sessions as it won't interrupt a blocking read, e.g. when the packet
reader is waiting for a packet.

Because that's what it does.

This can be used to combine e.g. the platform's SecureRandom
implementation with our own, so that a weakness in either source doesn't
harm security as long as the other source is strong.

We don't need two separate executors for long-running IO threads.