Remove Distribution::sample_iter

[sicking]

I might be misunderstanding the purpose of Distribution::sample_iter, in which case feel free to close this issue.

Best I can tell, the Distribution::sample_iter function is not useful. Implementations of the Distribution trait are required to still return a DistIter struct, which means that the only sensible implementation of Distribution::sample_iter is to use the default implementation.

So we might as well move the implementation of Distribution::sample_iter into Rng::sample_iter and get rid of Distribution::sample_iter.

This is arguably somewhat similar to issue 451, in that we currently have two ways of creating an iterator, Uniform::new(1, 30).sample_iter(rng) and rng.sample_iter(Uniform::new(1, 30)), whereas it seems having one would be better.

I'm happy to create a pull request if there's agreement.

[pitdicker]

In this case I like having both (both were at the same time actually). See #361.

[dhardy]

You mean not useful the way functions like Iterator::count are trivial to implement and have no sensible implementation besides the default? Or just because there are two functions for the same thing?

I agree with @pitdicker, in part because of symmetry with Rng::sample and Distribution::sample. It's weird for sure, but there's isn't really a good alternative I don't think (even the old Rand could be used in two ways).

[sicking]

Not useful in the sense that implementations of the trait can't provide a faster or otherwise more optimal implementation of the function.

But if the preference is to keep the function then so be it.