Rewrite NewRelicTrace to support fiber stops/starts

[rmosolgo]

Basically, using do ... end-type instrumentation can get confused when there's a Fiber.yield inside the block. (Dataloader does this when a field calls .load.) You can end up with situations where the sum of segments adds up to more than the total request time, because Fiber wait time is being added to the time when another Fiber is running. This makes sense-ish from New Relic's standpoint; they can't really know when the application handed over control in a way that "shouldn't count." See discussion at newrelic/newrelic-ruby-agent#3026.

So, what we can do is stop spans when there's a .yield called, then create a new span after the Fiber resumes control. I developed this in PerfettoTrace and I intend to roll it out to the other tracers, too.

---

I ran a little test app locally to confirm that this addresses the problem. Here's a before-after where the time tracking starts to line up:

TODO:

• Confirm on test NewRelic account
• Add tests demonstrating proper starts/stops
• Also do Accept Distributed Trace Headers From New Relic #5078 Still not clear on this