Fixes #4434.

This proposal replaces the previous CL 6855110. Due to issue 599, 64-bit atomic operations should probably be avoided, so use a sync.Mutex instead.

Benchmark comparisons against 025b9d070a85 on linux/386:

CL 6855110:

benchmark                        old ns/op    new ns/op    delta
BenchmarkTCPOneShot                 710024       727409   +2.45%
BenchmarkTCPOneShotTimeout          758178       768620   +1.38%
BenchmarkTCPPersistent              223464       228058   +2.06%
BenchmarkTCPPersistentTimeout       234494       242600   +3.46%

This proposal:

benchmark                        old ns/op    new ns/op    delta
BenchmarkTCPOneShot                 710024       718492   +1.19%
BenchmarkTCPOneShotTimeout          758178       748783   -1.24%
BenchmarkTCPPersistent              223464       227628   +1.86%
BenchmarkTCPPersistentTimeout       234494       238321   +1.63%

R=rsc, dvyukov, mikioh.mikioh, alex.brainman, bradfitz
CC=golang-dev, remyoudompheng
https://golang.org/cl/6866050