Turn down repeat count on ref counting and memory pool benches

It seems to me that there are many benches we run for several tens of ms when running them for 5-10ms would give us plenty of information.

These are the easiest two targets, and a spur to discussion.

I'd argue that on our target desktop platforms no 8888 bench should be taking longer than 20ms.

(This runs into a problem with morph, where even a single operation with a 10px kernel takes > 30ms - but do we really need to be testing blends that large?)

[bsalomon, 2012-09-12 12:44:20]

On 2012/09/12 11:30:24, TomH wrote:
LGTM

[bsalomon, 2012-09-12 12:44:44]

On 2012/09/12 12:44:20, bsalomon wrote:
> On 2012/09/12 11:30:24, TomH wrote:
> LGTM

And agreed in general about reducing bench times.

[TomH, 2012-09-13 15:42:57]

Committed in r5524. And committed to further reduction.