Remove extraneous Allocations

If we make a copy of the input `x` first, we can then operate on that copy in-place using views.
I think the code I have written plays it too safe with the write-accesses to `x` - by creating the needed parts as new zero-initialized-arrays.

We may get away with _not_ doing that and gaining some sizable speedups.

Locate the FFT implementation and trace where the input x is copied and where zero-initialized arrays are created for write access. Measure the current allocations and runtime first, then assess whether in-place views on a copy preserve correctness while reducing allocations. Done means the implementation remains correct and shows the intended speedup.

Written by the indexing model from the issue text.