We will need to return TransportErrors from Secure Session methods on
failures to extend the ErrorKind enumeration to actually contain the
transport error inside its ErrorKind::SessionTransportError variant.

This change has a number of implications. First of all, TransportError
is not copyable. Therefore we have to let go of the Clone and Copy
implementations on the ErrorKind as well as the Clone impl on Error.
One should not usually copy errors anyway so this should be okay.
Note that Error::kind() now returns a reference to the stored error
kind instead of its copy.

Another important thing is that TransportError does not implement
PartialEq as it is not possible to compare abstract errors in any
meaningful way. This implies that we cannot automatically derive
implementation of PartialEq for ErrorKind. We have to implement it
manually. We need to compare *kinds* of errors here so it is okay
to simply ignore the details of transport errors and treat them
as equivalent.

Now that we are able to store TransportError inside of a themis::Error
we can actually forward the error from the callback to the method call.
With this the user can actually see why the transport layer has failed.

Use the provided SecureSessionContext to temporarily store the error
while the control is still in the C code. Then we extract the error
and return it.

Such implementation does not support concurrent usage of Secure Session
from multiple threads but it's not thread-safe anyway so it is fine to
use this approach.

We can use negative return values to indicate implementation-specific
errors in Secure Session callbacks. The C code returns them as is for
us to inspect. While we're here, stop using bare constant "-1" and give
it a name. Also, handle the overflow with a different error kind.

There's another secure_session_receive() call here in this method. It
too can return special error codes.