Support `Clone` elements in `stack` and `select`

[bluss]

Related issue to #249

[jturner314]

This also applies to the .select() method (which is implemented in terms of stack).

[jturner314]

To avoid uninitialized data issues, probably the simplest implementation would be to choose the iteration order and memory layout such that a Vec could be extended with the data from each array in sequence, and then re-interpret the Vec as an Array at the end.

For example, let's say you want to stack arrays a and b along axis 2. I think this will work:

1. Permute the axes of a and b to move axis 2 to be the first axis, but preserve the ordering of the other axes.
2. Create a new Vec.
3. Extend the Vec from a.iter().cloned(), and then extend it with b.iter().cloned().
4. Convert the Vec into an Array. (The first axis of this array will correspond to axis 2 of a and b, and the other axes will be in order.)
5. Permute the axes of the Array to move the first axis to be axis 2.

This approach isn't necessarily the fastest because the iteration order may not be ideal, but it's relatively straightforward and doesn't need any unsafe code.

[gussmith23]

Wondering if there's any movement on this -- I think this is blocking me. I'm using ndarray to hold tensors of non-"simple" values (i.e. things that don't implement Copy, like other tensors!). I know that use-case probably raises hairs on some peoples' necks, but ndarray has been handling it pretty well up til now! It would be a bummer if this turned out to be a permanent block.

[gussmith23]

Also FWIW, I could take a crack at implementing something!

[bluss]

Let's get the #735 merged first (or base the work on that).

Non-copy elements aren't all that simple, but jturner has a plan here, and it sounds like it's a good one.

Non-copy elements have this trouble elsewhere: #797 (comment)