Bring those uses more in line with the pattern used in MLIR, i.e.
pack the cl::opts into a struct. We avoid ManagedStatic entirely by then
instantiating that struct as a static function variable.
Remove additional uses of ManagedStatic in the affected files at the
same time.
Is the problem described by this comment not an issue anymore? I guess the timing of the call to initTimerOptions is more predictable since we aren't using a global variable, but is it guaranteed to happen before the StatisticInfo constructor runs?