Use range based zeroing for >= Go 1.5

[jrick]

Go 1.5 optimizes code such as the following:

for i := range slice {
        slice[i] = 0
}

to a duff's device to zero memory in the slice. With Go tip, it benchmarks faster than the zero.Bytes function.

BenchmarkRange32    300000000            4.14 ns/op
BenchmarkRange64    300000000            4.26 ns/op
BenchmarkBytes32    200000000            7.91 ns/op
BenchmarkBytes64    100000000           13.6 ns/op

The range-based zero should be enabled for all versions but Go 1.4 using build tags. When Go 1.5 is released, we will stop supporting Go 1.3, so only 1.4 needs the custom implementation. After 1.6 is out, the custom implementation can be removed entirely.