context:

after raising socket pool size from 1 to 5 we observed a dramatic spike in false negative healthchecks
it is still unclear why, but one possible explanation is that because messages are being sent faster, we are sending healthchecks before we have a chance to register their callback, thus the healthcheck response arrives before its callback handler has been registered, and the callback handler, once registered, waits for a healthcheck response that never arrives (because it arrived before the handler started listening)

changes:

this MR tests the above hypothesis by registering healtcheck callbacks before sending them. if it works, we should observe a decrease to close to zero false negative healthchecks