Investigate performance impact of flat image buffers (Vec<f32>) vs planar buffers ([Vec<f32>; 3])

[FreezyLemon]

Potential option for #11.

For every scale, we first convert to Xyb (which uses the same data: Vec<[f32; 3]> as LinearRgb) and then to a planar data representation. Instead of a Vec<[f32; 3]>, we get three Vec<f32>, both with len == amount_of_pixels.

The conversion between these two representations is probably not that important in terms of speed (although it does require three fairly large allocations), I wonder if we could use a single Vec<f32> with len == 3*amount_of_pixels instead.

The naive thought I had was to just inline all planes after each other (&data[0..plane_size] is the X plane, for example). This would still require a new allocation, but might be faster when blurring or doing the metric calculations.

Another thought I just had: Maybe we could even reuse the Vec<[f32; 3]>'s allocation. Casting to Vec<f32> is relatively simple with some unsafe code (we already do it in yuvxyb). Bit difficult to tell what performance implications this would have though.