Low async throughput under heavy load when appender is configured to *not* drop events

[brenuart]

The Async appenders show very bad performance when configured to not drop events when the ring buffer is full (appendTimeout=-1).
The async appender first try publish the event in the ring buffer then sleeps for appendRetryFrequency milliseconds before retrying. This relatively long sleep affects performance dramatically.

An alternative is to use an exponential backoff strategy capped by the appendRetryFrequency. This should give better performance while still limiting the CPU consumption.

[brenuart]

Another problem arise in CPU constraint environments when more threads than the available cores are retrying at the same time. The first retries happen at high rate (a few nanoseconds pauses) then slow down up to the configured appendRetryTimeout. This behaviour increases throughput at the expense of additional CPU cycles. We should limit the retries to a single thread at once to avoid overwhelming the system when only a few cores are available.