The improvement is obtained by eliminating the zero
initialization of a large structure that is only
needed when the fast path fails.

Also add a missing roundtrip test for float32s.

benchmark                              old ns/op    new ns/op    delta
BenchmarkAppendFloatDecimal                  301          180  -40.20%
BenchmarkAppendFloat                         486          388  -20.16%
BenchmarkAppendFloatExp                      492          383  -22.15%
BenchmarkAppendFloatNegExp                   478          370  -22.59%
BenchmarkAppendFloatBig                      650          541  -16.77%
BenchmarkAppendFloat32Integer                308          180  -41.56%
BenchmarkAppendFloat32ExactFraction          449          333  -25.84%
BenchmarkAppendFloat32Point                  494          390  -21.05%
BenchmarkAppendFloat32Exp                    488          387  -20.70%
BenchmarkAppendFloat32NegExp                 488          378  -22.54%

R=r, rsc
CC=golang-dev, remy
https://golang.org/cl/6346081