Benchmark specialized distribution implementations

[vks]

As noted in #1062, this demonstrates that Exp1 is slower than Exp, which kind of defeats its purpose.

[dhardy]

Presumably the optimiser can remove the addition/multiplation by identity here.

There is still one other potential reason to use Exp1/StandardNormal: to reduce data size within some other sampler instanced many times. This doesn't require explicit documentation, except perhaps pointing out that they are the building blocks.

[dhardy]

I guess we might as well merge this.