multiscales support for `labels`

[HelenaLC]

• following implementation of multiscales for images
• if we see repetitive usage of the same code/functions etc., might be work considering defining a class union (say, sdArray) to be used internally for defining shared S4 methods (same goes for sdFrame if we see lots of similar behavior for point and shape layer)

[melonora]

Maybe helpful to know that in python SpatialData we have RasterSchema which is used for both the ImageModels and the LabelModels (both 2D and 3D). Pretty much the only difference between the 2 is the dimensions they support (labels never have a c dimension). These models parse the data in an Xarray DataArray in case of single scale or DataTree in case of multiscale.

[Artur-man]

Quick question: what is exactly the difference between defining class unions, and defining methods for multiple classes like below:

SpatialData/R/mask.R

Lines 88 to 92 in 18a25d6

	setMethod(".mask", c("ImageArray", "LabelArray"), \(a, b, fun=mean) {
	# TODO: somehow rewrite w/o realizing everything
	# at once (maybe w/ 'DelayedArray::blockApply'?)
	.a2v <- \(.) as.vector(as.array.Array(.))
	stopifnot(dim(a)[-1] == dim(b))

[HelenaLC]

The example you linked here does not define a method for multiple classes, but for a specific class signature, where a is an ImageArray and b is a LabelArray; any other classes or different ordering will not dispatch the method.

Per my understanding, class unions allow defining (once) methods that can be the same across all subclasses, which might save redundant implementation is some cases... E.g., labels and images might share a set of methods (through a class union), but also have class-specific methods (for their own class only).

An example of this is SpatialDataElement, which I allows for shared methods across layers, e.g., to extract Zattrs.

[Artur-man]

Ok, this is done and merged in #116. If we come across to some other problems we can discuss again.