Improve performance of the metrics collection

The current metrics system add a 15% slowdown in the events processing of the pipeline. We should improve that situation. I believe we have some room to grow to improve the situation. I think is reasonable to improve the performance by 5%.

Lets try the following:
- We have some allocation required for the lookup cache which I believe could be removed.
- We are using the atomicfixnum type in the metric store, since we have multiples threads updating the same values we should look at `LongAdder` to see if we can some performance improvements #4508
- Profile the code to see the bottlenecks.

Followup of #4572