Implement inplace `circshift`

[musm]

We have circshift but no inplace version, which would be very useful. There is a rough implementation in the lbm3d threads test.

[stevengj]

(See also the discussion at the end of #9822.)

[stevengj]

@Musmo, the circshift you're referring to is based on a chase-the-cycles approach. The Knuth algorithm (based on two reversal passes) is far better for arbitrary shifts.

[stevengj]

For example, here is a working 1d circshift!:

function circshift!(a::AbstractVector, shift::Integer)
    n = length(a)
    s = mod(shift, n)
    s == 0 && return a
    reverse!(a, 1, s)
    reverse!(a, s+1, n)
    reverse!(a)
end

The only thing I'm not sure of is how to generalize this to arbitrary dimensions without writing out a bunch of explicit loops for 2d, 3d, etcetera arrays. @timholy, is this something that your cartesian iterators could handle?

(Also, it would be nice to fuse the reversal loops in higher dimensions so that only 2 passes, not 2k passes, are required for shifting a k-dimensional array along all k dimensions simultaneously. But maybe in practice people usually shift only along a 1 or 2 dimensions at a time?)

[stevengj]

@timholy, is it worth keeping this open for a truly in-place circshift!, as opposed to a circshift! with pre-allocated output?

[timholy]

I'm fine with that, if others want that. We don't have a truly in-place transpose!, either. (Not that I want one, but someone might.)

So feel free to reopen if you are interested.