Currently, trace processing interleaves state/statistics updates and
emitting trace viewer objects. As a result, if events are being
filtered, either by time or by goroutines, we'll miss those
state/statistics updates. At best, this leads to bad statistics;
however, since we're now strictly checking G state transitions, it
usually leads to a failure to process the trace if there is any
filtering.
Fix this by separating state updates from emitting trace object. State
updates are done before filtering, so we always have correct state
information and statistics. Trace objects are only emitted if we pass
the filter. To determine when we need to emit trace counters, rather
than duplicating the knowledge of which events might modify
statistics, we keep track of the previously emitted counters and emit
a trace counter object whenever these have changed.