The algorithm is the same as in the double-conversion library
which also implements Florian Loitsch's fast printing algorithm.
It uses extended floats with a 64-bit mantissa, but cannot give
an answer for all cases.

                           old ns/op  new ns/op  speedup
BenchmarkAtof64Decimal         332        322      1.0x
BenchmarkAtof64Float           385        373      1.0x
BenchmarkAtof64FloatExp       9777        419     23.3x
BenchmarkAtof64Big            3934        691      5.7x
BenchmarkAtof64RandomBits    34060        899     37.9x
BenchmarkAtof64RandomFloats   1329        680      2.0x

See F. Loitsch, ``Printing Floating-Point Numbers Quickly and
Accurately with Integers'', Proceedings of the ACM, 2010.

R=ality, rsc
CC=golang-dev, remy
https://golang.org/cl/5494068