Make Take and Skip iterators DoubleEnded when constructed from ExactSize

[ghost]

Fixes #12784

[ghost]

Skip could also be specialized for A: RandomAccessIterators + ExactSize to make it more efficient in that case.

[bill-myers]

Might not be a good idea to merge this, because it has unintuitively slow performance.

Specifically, you might expect linked_list.iter().take(K).rev() to take O(k) operations, but instead it takes O(n) where n is the size of a linked list (and in fact, the current implementation will be slow for all iterators).

Those wanting to do this can use iter().rev().skip(N - K) instead, which makes the performance impact obvious.

Or they can more sensibly decide to collect the first K elements in a vector, and then reverse that, which is the time-efficient way of doing it.

[ghost]

Agreed.

It'd be possible to only implement these for A: ExactSize + RandomAccessIterator and then take(k).rev() could indeed be O(k). As would skip(n).take(k), which currently is O(n + k).

[alexcrichton]

Travis has some failures which should get addressed before merging