Tracing is uneasy with only the public interface

[MichaelMure]

I wanted to monitor the messages dropped due to #398 but the tracing API make that difficult. I ended up writing the following (obviously not great) code:

type tracer struct {
    ingressDropped *stats.Int64Measure
    egressDropped  *stats.Int64Measure
}

func (t tracer) Trace(evt *pubsub_pb.TraceEvent) {
    switch evt.GetType() {
    case pubsub_pb.TraceEvent_REJECT_MESSAGE:
        if evt.RejectMessage == nil {
            return
        }
        switch evt.RejectMessage.GetReason() {
        case "validation queue full", "validation throttled":
            stats.Record(context.Background(), t.ingressDropped.M(1))
        }
    case pubsub_pb.TraceEvent_DROP_RPC:
        stats.Record(context.Background(), t.egressDropped.M(1))
    }
}

This code:

• use string constants copied from the package as they are not public
• require a deep understanding of the package internals to catch the correct case
• imply that we are doing an unnecessary marshaling to a protobuf message

The internalTracer interface looks better in my opinion but it's also not publicly accessible.

[vyzo]

So the right solution for this is to simply expose the lower level interface from internalTracer and export the reason strings. This should give you a more powerful hook that avoids the protobuf round trip.

[vyzo]

See #403.