SFTP connection cache should be removed

robey, I'd love if if you can tackle the first bit of this - making the stub server accept multiple connections.

here's a patch that changes SingleListener to SocketListener and keeps listening for new connections until explicitly asked to stop (via TestCase.tearDown).

this lets the sftp unit tests continue to pass, but that means nothing since they don't try to use multiple connections yet. :)

if i comment out the connection cache using this patch, one of the tests locks up pretty quickly. i haven't had a chance to look at that yet, but i will continue later.

Ah, found out what the problem was and was a simple fix. Since the loopback socket version is just handling sftp requests inline, and transports are never "closed", we can't wait around for the socket to be closed. So I just spin it off into a new thread each time.

With this patch, I can comment out the connection cache, and the unit tests still pass -- except for one which is trying to verify that the connection cache is working. :)

I believe the agreed upon fix is to be able to supply additional transport objects to 'get_transport()' and ask it to re-use them if it can.

This needs to propagate up the stack a bit (so Branch.open_containing() might take a list of transports to re-use).

But this makes the 'cache' a very localized operation, and not global througout bzrlib.

Patch available shortly, hope robert don't mind that I assigned it to myself ;-)