Reduce clones in RepeatN

[cuviper]

The standard library is optimized to avoid cloning when returning its last item. We can do the same for our version, and the savings are a lot greater because the waste was multiplied across every split instance as the parallel iterator ran. Now we avoid cloning when returning the last item of any split, so we don't hold any value at all when the count is zero.

[adamreichold]

LGTM

As an aside, I wonder whether for Rayon, a variant that requires T: Sync + Clone and packages the element in an Arc would be helpful to avoid the repeated dispatch on the ::Empty case.

[cuviper]

As an aside, I wonder whether for Rayon, a variant that requires T: Sync + Clone and packages the element in an Arc would be helpful to avoid the repeated dispatch on the ::Empty case.

I think if we add repeat_with, you could easily roll your own to that effect, like repeat_with(move || x.clone()), since that would use F: Fn + Sync which would work based on the captured Sync. I guess we might also want something like repeat_n_with and the same kind of take/zip methods we have on Repeat, so you can have an IndexParallelIterator of this.