Title: Support for request hedging (racing) on retry timeout

Description:
This issue is similar to #2784 (which discusses support for sending multiple simultaneous requests and using the first response to manage tail latency) however specifically pertains to handling of a timed out request. This would allow for sending a comparatively lower overall request volume than straight up hedging while being resilient to slow upstreams. It's essentially the same per try timeout behavior in place today, except the original request will not be canceled and the first response received can be used.

I imagine it would make sense to implement #2784 first since that seems like less of an architectural change to the router and then add options to only send the extra requests when the per try timeout is hit.

I'm interested in taking it on