In -test.short mode, skip measuring the upper bound of time
sleeps. The API only guarantees minimum bounds on sleeps,
anyway, so this isn't a bug we're ignoring as much as it is
simply observing bad builder virtualization and/or loaded
machines.

We keep the test in full mode where developers will
presumably be running on a lightly-loaded, native, fast
machine.

R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/5713044