Uses atomic memory accesses to avoid the need to acquire
and release schedlock on fast paths.

benchmark                            old ns/op    new ns/op    delta
runtime_test.BenchmarkSyscall               73           31  -56.63%
runtime_test.BenchmarkSyscall-2            538           74  -86.23%
runtime_test.BenchmarkSyscall-3            508          103  -79.72%
runtime_test.BenchmarkSyscall-4            721           97  -86.52%
runtime_test.BenchmarkSyscallWork          920          873   -5.11%
runtime_test.BenchmarkSyscallWork-2        516          481   -6.78%
runtime_test.BenchmarkSyscallWork-3        550          343  -37.64%
runtime_test.BenchmarkSyscallWork-4        632          263  -58.39%

(Intel Core i7 L640 2.13 GHz-based Lenovo X201s)

Reduced a less artificial server benchmark
from 11.5r 12.0u 8.0s to 8.3r 9.1u 1.0s.

R=dvyukov, r, bradfitz, r, iant, iant
CC=golang-dev
https://golang.org/cl/4723042