Currently trace.lock is reentrant in a few cases. AFAICT, this was
necessary a long time ago when the trace reader would goparkunlock, and
might flush a trace buffer while parking the goroutine. Today, that's no
longer true, since that always happens without the trace.lock held.
However, traceReadCPU does still rely on this behavior, since it could
get called either with trace.lock held, or without it held. The silver
lining here is that it doesn't *need* trace.lock to be held, so the
trace reader can just drop the lock to call traceReadCPU (this is
probably also nice for letting other goroutines flush while the trace
reader is reading from the CPU log).