WIP: Support slices like '3:10:mean(2)'

[danielballan]

Closes #32

As proposed in #32, this will support:

• Reducing an N-dimensional array to an array of lower dimension by averaging over one or more dimensions. For example, average an image time series like ?slice=::mean
• Downsamplig an N-dimensional array over some axes. For example, downsampling an image time series by a factor of 2 in x and y: ?slice=:,::mean(2),::mean(2)

The downsampling with employ local averaging as in https://scikit-image.org/docs/stable/api/skimage.transform.html#downscale-local-mean

Incidentally this removes the use of eval or client input, which I like. it was technically safe because it was guarded by regex, but it still made me nervous.

[EliotGann]

This would be extremely helpful for my igor Tiled client, where viewing downsampled images is really important - saving client memory and working with slower internet speeds. It also will allow for trivial data reduction for already resampled data. Essentially I will be able to construct a URL for users to pull valid 1D scattering curves from 2D datasets.

[gwbischof]

This should be rebased.
The test should be updated to test downsampling and averaging.
Client should be updated to support these updates.

[danielballan]

This related project shows another way to spell this which I think I like better:

Leave the slice syntax as is, taking standard numpy-style slices. Add an optional query parameter downsample_method. Its default value can be stride but it can accept other values mean, max, min….

https://google.github.io/tensorstore/driver/downsample/index.html

[danielballan]

It occurred to me later downsample_method as implemented there assumes we are using the same downsample method for all axes, whereas the original approach we considered here, ?slice=:,::mean(2),::mean(2), supports a more general case of different downsample methods on different axes.

I suppose we could combine the two and support:

?downsample_method=mean  # same on all axes
?downsample_method=mean,mean,max  # different for each axis

but that creates a dependency with the value of slice that seems a bit odd...I might be back to preferring our original plan.