fix a race condition in PipeListener

[natefinch]

There was an obvious race between PipeListener's Accept and Close modifying l.closed, found with go test -race, and a less obvious race that if someone calls Close before we call connectnamedpipe, the connect would hang forever. Now we have a lock over the whole of Accept and Close (except while doing WaitSingleObject, which will corrently return an error when we Close), so Close can't screw up connect named pipe.

[natefinch]

Testing done: there was a test in juju (under worker/metrics/sender) that would occasionally hang (approx 1/10 times)... with this fix, no more hang in 300 trials.

(also go test -race used to find a race condition and now does not)