Low priority to come back to this later. The ULT that runs sparkline data collection needs to idle most of the time, which seems like a sensible use of margo_thread_sleep(). There is a bug that prevents that from working correctly that we need to dig into.

Right now the code is using ABT_thread_yield() instead, which could result in the sparkline ULT being scheduled too frequently in some cases.

This issue is not relevant in master until the breadcrumb MR from @sramesh is merged.