Feature(new_metrics): implement the percentile

[empiredan]

The percentile is a metric type that samples observations as is described in #922. The size of samples has an upper bound. Once the maximum size is reached, the earliest observations will be overwritten.

On the other hand, kth percentiles, such as P50, P90, P95, P99, P999, will be calculated periodically over all samples. The kth percentiles which are calculated are configurable provided that they are of valid kth_percentile_type (i.e. in kAllKthPercentileTypes).

The most common usage of percentile is latency, such as server-level and replica-level latencies. For example, if P99 latency is 10 ms, it means the latencies of 99% requests are less than 10 ms.

The percentile is implemented by the finder for nth elements. Each kth percentile is firstly converted to nth index; then, find the element corresponding to the nth index.