GPU support for ndfilters & imread modules

[GenevieveBuckley]

This branch contains broader support for GPU array computation with cupy & dask-image.

Related PR: #150
Related issue: #133

TODO

• ndfilters - done! (Note: @jakirkham used a clever way to avoid repetition with the minimum, median, and maximum filters, but this PR opts for a small amount of repetition for the sake of simplicity - it seemed easier to keep the docstring wrapper function and the dispatching working well this way)
• imread - done!
• Tests explicitly using cupy arrays (even if we have to mark them as skipped on CI)

For future PRs

• GPU support for ndmorph - easy to implement (see gist here), but it relies on having cupy support for the functions binary_dilation() and binary_erosion() which does not currently exist.
• GPU support for ndfourier - problem: lots of numpy specific array creation in dask_image.ndfourier._utils
• GPU support for ndmeasure - problem: lots of numpy specific array creation in distributed labelling. I still need to look at what needs to be done to support all the ndmeasure functions.

[GenevieveBuckley]

### Request for help: lazy dispatch for correlate.
I'm at my wits end here, so if you have any suggestions @jakirkham I'd be happy to hear them.

EDIT: I have found the place where I made my typo so I no longer need help here, hooray!

What works:

• Dispatch for convolve with lazy cupy function registration
  • The implementation for convolve is virtually identical to the implementation for correlate, I don't understand why they are behaving differently.
  • All the other ndfilter function dispatchers seem to work fine too, it's just correlate that's not working
• The non-lazy version of the correlate dispatch for cupy arrays also works. You can try this yourself by commenting lines 66-67 in dask-image/dask_image/dispatch/_dispatch_ndfilters.py and then indenting lines 68-73 below that.

What doesn't work:

• Dispatch for correlate with lazy cupy function registration

EDITED TO ADD: What the problem actually was that I accidentally decorated two functions with @dispatch_correlate.register_lazy("cupy"). Fixed it in commit 3eb3ec2

[chrisroat]

Regarding binary_dilation and binary_erosion for ndmorph, there is cupy support through the following repo. Not sure if you can borrow or depend on it:
https://github.com/mritools/cupyimg

[GenevieveBuckley]

Regarding binary_dilation and binary_erosion for ndmorph, there is cupy support through the following repo. Not sure if you can borrow or depend on it:
https://github.com/mritools/cupyimg

Good to know, thanks @chrisroat

[GenevieveBuckley]

Regarding binary_dilation and binary_erosion for ndmorph, there is cupy support through the following repo. Not sure if you can borrow or depend on it:
https://github.com/mritools/cupyimg

Good to know, thanks @chrisroat