Correct cleanup on repeated waiting for never arriving messages

[ThomasDebrunner]

When calling expect(..) a shared_ptr to a std::promise is returned.
Thie shared pointer is being held by both the mav::Connection as well as the caller. The state is being cleared in the connection, iff 1) a message arrives that fulfils the expectations or 2) there is an exception on the network.

Now, if the caller excepts something that will never occur, the expectation is being held onto by the connection forever, even when the caller has long forgotten about it.
The typical case is that the caller has received a timed out while waiting for, or the expectations has gone out of scope for some other reason. In this case, this could be a resource leak.

This PR changes the internal accounting in the connection to a std::weak_ptr. If any message is received, and the handler stumbles upon an expired weak pointer, it is cleared from the callback map.

This way we make sure that we only hold references to promises that the caller is still keeping track of.

[eyeam3]

I had the test app running for about 10 minutes and the CPU load stayed constant.