Roughly 33% faster for simple cases, probably more for complex ones.

Before:

mallocs per Sprintf(""): 4
mallocs per Sprintf("xxx"): 6
mallocs per Sprintf("%x"): 10
mallocs per Sprintf("%x %x"): 12

Now:

mallocs per Sprintf(""): 2
mallocs per Sprintf("xxx"): 3
mallocs per Sprintf("%x"): 5
mallocs per Sprintf("%x %x"): 7

Speed improves because of avoiding mallocs and also by sharing a bytes.Buffer
between print.go and format.go rather than copying the data back after each
printed item.

Before:

fmt_test.BenchmarkSprintfEmpty	1000000	      1346 ns/op
fmt_test.BenchmarkSprintfString	500000	      3461 ns/op
fmt_test.BenchmarkSprintfInt	500000	      3671 ns/op

Now:

fmt_test.BenchmarkSprintfEmpty	 2000000	       995 ns/op
fmt_test.BenchmarkSprintfString	 1000000	      2745 ns/op
fmt_test.BenchmarkSprintfInt	 1000000	      2391 ns/op
fmt_test.BenchmarkSprintfIntInt	  500000	      3751 ns/op

I believe there is more to get but this is a good milestone.

R=rsc
CC=golang-dev, hong
https://golang.org/cl/166076