If the chosen backend is sick or fails to give us a connection,
we rescale the keys position in the backends window onto the
all the remaining backends.

The upside of this is that we get a stable and predicatable
hash distribution.

The downside of this is that you are very unlikely to spot the
healthy .weight=1 backend between two .weight=100 sick ones before
your retries run out.

Inspired by: patch from WaveCDN