Related to #1423 and #1570.

We should consider further enhancements to this feature:

• When using /healthcheck/fail admin endpoint, have some period of drain time before starting 100% drain/close.
• Add ability for header value to specify things like TTL for failure if used for load shedding. (Need to think about whether this counts as an outlier ejection vs. active health check).
• Potentially ASSERT more thread locality stuff.

cc @alyssawilk