WIP: test peer churn under sustained forwarding load

[jcorbin]

Finally reproduces the production failure that we've been hunting, TODO:

• find and fix the cause of the zombie connections to the turnips
• understand all the newly provoked logs, fix, or adapt test expectations as appropriate

[jcorbin]

it's also amusing to observe that the endpoint delay that I added will probably allow to test at higher scale of other parameters, even in travis, since the cpu is less contended due to the slow endpoint causing the test machine to smear over time more.