need to be more aggressive about transport thread safety

If you run serve-branches http://whatever and make simultaneous or very nearly so requests to loggerhead, things explode in exciting ways. This is because http transport share connections across clone() calls, so although each thread serving a request gets its own clone()d transport they share a connection and using this connection from more than one thread at once doesn't work at all well (a call may get bytes that are the response to another thread's call, for example).

Probably what we should do is have a threading.local object that stores a transport for each thread (it's what we do on Launchpad and it works well there).