Collect container pull duration metric

[Nevon]

In order to monitor and optimize container startup times, it would be very helpful if the Datadog container integration collected the pull time duration and exposed it as a metric.

As an example of prior art, kubelet collects this as a histogram bucketed by image size which might be a good way to avoid tag cardinality explosion while still allowing this metric to be usefully compared (it doesn't help to monitor pull times if the reason they go up is because the images get larger).