Use atomic.LoadUint32(&done) instead of
atomic.AddInt32(&done, 0) on fast path.

benchmark            old ns/op    new ns/op    delta
BenchmarkOnce            13.40         7.26  -45.82%
BenchmarkOnce-2          22.90         4.04  -82.36%
BenchmarkOnce-4          25.60         2.16  -91.56%
BenchmarkOnce-8          25.80         1.38  -94.65%
BenchmarkOnce-16         24.40         1.33  -94.55%

(on HP Z600, 2 x Intel Xeon E5620, 8 HT cores, 2.4 GHz)

R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/4787041