Added bench to test multiple readPixels case

I think this captures the SchemerDreamer use case:

running bench [640 480]  readpix  8888: cmsecs =   0.62   GPU: cmsecs = 103.74 gmsecs = 104.05

[bsalomon, 2012-09-04 13:21:48]

On 2012/09/04 13:18:31, robertphillips wrote:

LGTM, should we have a variant that does a trivial draw operation between each
iteration so that we can see the relative gain of a readback cache?

[robertphillips, 2012-09-04 13:41:40]

I actually already have a straw-man fix for this but I wanted to deliver it
separately so we would have before and after times. I don't know if we would get
any benefit from always benching the degenerate write-read-loop case.

committed as r5381