prevent devious scheduling from reporting negative size to kamon

[schlawg]

ConcurrentLinkedQueue.add is not atomic with respect to AtomicInteger.getAndIncrement. I think we should increment the size variable before we call add to address this race condition:

Thread A is executing the flush() loop.
Thread B calls flushQ.add, the channel is added, but the thread is interrupted before incrementing size.
Thread A empties the queue and now flushQ.size is actually -1.
Thread A takes a 1ms nap while thread B remains suspended.
Thread A wakes up and enters flush() again, assigning flushQ.estimateSize() (which is still -1) to qSize
Thread B finally resumes and corrects the size variable, but the spurious negative value has already been captured and will be reported to kamon.

A negative could spike the y-axis scale if grafana or another process in the reporting chain is configured to expect unsigned values.

[schlawg]

The PR was inspired by this graph