Further questions about truncation error reporting for `svd_trunc!`

[kshyatt]

Follow up to #102 and #113 (comment).

We have some outstanding issues with handling optional return of the truncation error for svd_trunc and friends.

These are:

• Should epsilon always be passed by reference, as an array or Ref? This would allow us to pass it as a reference to truncation_error!
• How to handle the compute_error keyword argument in svd_trunc (without a !) for algorithm selection?

[lkdvos]

• I think passing it by Ref might be annoying for the GPU (if I understand that correctly?), so I would say maybe to use a 0-dim array?

• Is there a way to put the compute_error keyword in the algorithm structs? At least that puts this keyword on the same footing as all other keywords and we don't have to handle that one in a different way

[kshyatt]

• There's the noble RefOrCuRef that CUDA has, there might be an AbstractRef we could use as well. I'll look into it.

• Would it make sense to put it in the TruncationAlgorithm wrapper struct, since that's the only place it would be needed?

[kshyatt]

A follow-up question:

• Should I be able to pass an array for epsilon to svd_trunc? (without the !)

[kshyatt]

I think we can close this for now with the addition of the _no_error(!) functions