Allow cutoff predictions, support empty background pixels and tests

[andsild]

Hi. Let me know if you want me to try and break down the PR into smaller PRs.

Otherwise, this PR is a first step to work with AML in DSA.
Changes are 1:

1. Adds an option for cutoff predictions to be made, so that you only do predictions on n unlabeled features per round/epoch. All labeled samples will always be included. This is to speed up training in cases where you have many slides with many annotations (in our AML case, several millions). The benefit is speed, the downside may be that feature files no longer contain all the data if a user wants to download them.
2. Define a background superpixel in cases for sparse pixelmaps (in my case, white blood cells where we are not interested in the space between them). The background superpixel will always be ignored. This is mostly for the UI (histomics-label). I tried a cleaner approach where we did not use dummy/mock values for this, but it requires changes to several parts of the code, also in histomics-label (to index at 1) large_image (for drawing superpixels where the indexing starts at 1) and histomicstk(to generate superpixels with starting value 1). I wanted to start simple by allowing an empty, unlabeled background pixel of size 1, which makes it more or less invisible in the UI.
3. Sort all labeled superpixels last in the "confidence" order from predictions to prevent them from re-emerging first to the user in the filmstrip during active learning.
4. Added tests to make sure my code (hopefully) doesn't break anything. The tests can also be used to benchmark the code.
5. Using CUDA is now an optional parameter. I've mostly done this to make tests run on the CPU only. But users can now toggle GPU/CPU on and off in slicer.
6. Miscellaneous, spelling fix/print statements

I have tested in DSA with my own AML slides and also using the default superpixels from the UI. The tests should also help to establish a baseline of what works and doesn't (test_full_predict starts from scratch with a slide of MNIST numbers and gets an accuracy around 80%`).

For the UI, see also DigitalSlideArchive/histomics-label#207

[andsild]

this should be true by default

[andsild]

training only uses labeled samples

[manthey]

@andsild I find that having multiple PRs accelerates review -- we can get simple things merged in and then the complex things look simpler. I often will do a git diff master > some_diff_file.txt, then edit the diff file to have the isolated change and apply it to a new branch based on master to tease apart the PRs (but if you use an IDE it might do all that more conveniently).

[andsild]

@andsild I find that having multiple PRs accelerates review -- we can get simple things merged in and then the complex things look simpler. I often will do a git diff master > some_diff_file.txt, then edit the diff file to have the isolated change and apply it to a new branch based on master to tease apart the PRs (but if you use an IDE it might do all that more conveniently).

I get that 😄 good practice to do anyway

I've opened up

• Misc changes #30
• Make using CUDA an optional parameter #31
• Support cutoff number of features per slide #32
• Allow a background pixel of [0,0,1,1], which will always be ignored #33

which is just this PR divided into multiple others