Title: uneven request distribution when using weighted LR LB

Description:

When LR LB is used with weights, each upstream tends to send a large series of requests to one host before moving onto the next. For example, we've seen 50 sequential requests into a single box when using a weight of 50. It seems like the LB strategy is to send 'n' requests to a given host before moving onto the next one.

Repro steps:

Setup a weighted cluster and observe the request distribution into