This old __per_cpu_data define wasn't enough if an arch wants to
use the gcc __thread prefix (thread local storage), which needs to
go *before* the type in the definition.  So we have to go for a
DECLARE macro, and while we're there, separate DECLARE and DEFINE,
as definitions of per-cpu data cannot live in modules.  This also
means that accidental direct references to per-cpu variables will
be caught at compile time.