Improve in accounting/metric memory usage and throughput

As seen in a recent CBSE, if a client hits the n1ql stats hard enough, we could have the GC struggle to clear slices allocated to display request stats.

Clean up the metrics/sample type not to make copies of heaps every time an aggregate is requested (and the /admin/stats endpoint requests a fair few for each call), rejig the interface to allow to pass a timestamp (rather than calling several times time.Now() with the same value), and improve on contention while inserting new samples.