The speedup is a combination of unrolling/specializing
the actual code and also making the compiler generate better code.

Go 1.0.1 (size: 1239 code + 320 data = 1559 total)
md5.BenchmarkHash1K   1000000	   7178 ns/op	 142.64 MB/s
md5.BenchmarkHash8K    200000	  56834 ns/op	 144.14 MB/s

Partial unroll  (size: 1115 code + 256 data = 1371 total)
md5.BenchmarkHash1K   5000000	   2513 ns/op	 407.37 MB/s
md5.BenchmarkHash8K    500000	  19406 ns/op	 422.13 MB/s

Complete unroll  (size: 1900 code + 0 data = 1900 code)
md5.BenchmarkHash1K   5000000	   2442 ns/op	 419.18 MB/s
md5.BenchmarkHash8K    500000	  18957 ns/op	 432.13 MB/s

Comparing Go 1.0.1 and the complete unroll (this CL):

benchmark               old MB/s     new MB/s  speedup
md5.BenchmarkHash1K       142.64       419.18    2.94x
md5.BenchmarkHash8K       144.14       432.13    3.00x

On the same machine, 'openssl speed md5' reports 441 MB/s
and 531 MB/s for our two cases, so this CL is at 90% and 80% of
those speeds, which is at least in the right ballpark.
OpenSSL is using carefully engineered assembly, so we are
unlikely to catch up completely.

Measurements on a Mid-2010 MacPro5,1.

R=golang-dev, bradfitz, agl
CC=golang-dev
https://golang.org/cl/6220046