PERF: avoiding an array copy for single-pass convolution

[neutrinoceros]

The minimal amount of array allocations when looping is 1 for any number of passes, except for 1, where it's 0.
Given that this is expected to be a very common use case, it seems worth trying to special case it. However:

• I'd need a way to change the internal convolve routine's signature such that it accepts an Array2 or an ArrayView2. Currently, I don't know how to do that
• in my attempt to quantify the overhead of this unecessary allocation, I estimated that it was at most 10% of the total time needed for a pass, so it may not be worth the complexity

[neutrinoceros]

Actually it's possible that no copy is ever needed. I'm thinking iterative LIC should be equivalent to a single pass with a different (and longer) kernel. I need to work out the details but if my hunch is correct I should be able to dramatically reduce the time needed for large iteration numbers.

[neutrinoceros]

I thought of this idea some more and concluded that it wasn't a good idea:

• Using a longer kernel is not, in general, equivalent to multi-iterations processes (because starting points are reset at each iteration)
• there's an infinite amount of ways to modify a user provided kernel on the fly, none of which are obviously correct, and all of which would be suprising in my opinion.

Overall, it wouldn't add any actual value.

The original ask (avoiding a copy for single-step iterations) still stands though.

[neutrinoceros]

I still want to explore this idea, but I'll note now that if binaries end up doubling in size to support this, I will probably close the issue as not-planned