Currently the benchmarks lie to testing package by doing O(N)
work under StopTimer. And that hidden O(N) actually consitutes
the bulk of benchmark work (e.g includes GC per iteration).
This behavior accounts for windows-amd64-race builder hangs.

Before:
BenchmarkReaderCopyOptimal-4	 1000000	      1861 ns/op
BenchmarkReaderCopyUnoptimal-4	  500000	      3327 ns/op
BenchmarkReaderCopyNoWriteTo-4	   50000	     34549 ns/op
BenchmarkWriterCopyOptimal-4	  100000	     16849 ns/op
BenchmarkWriterCopyUnoptimal-4	  500000	      3126 ns/op
BenchmarkWriterCopyNoReadFrom-4	   50000	     34609 ns/op
ok  	bufio	65.273s

After:
BenchmarkReaderCopyOptimal-4	10000000	       172 ns/op
BenchmarkReaderCopyUnoptimal-4	10000000	       267 ns/op
BenchmarkReaderCopyNoWriteTo-4	  100000	     22905 ns/op
BenchmarkWriterCopyOptimal-4	10000000	       170 ns/op
BenchmarkWriterCopyUnoptimal-4	10000000	       226 ns/op
BenchmarkWriterCopyNoReadFrom-4	  100000	     20575 ns/op
ok  	bufio	14.074s

Note the change in total time.

LGTM=alex.brainman, rsc
R=golang-codereviews, alex.brainman, rsc
CC=golang-codereviews
https://golang.org/cl/51360046