Register deterministic tokenization for range objects

[jrbourbeau]

Currently our tokenization of range objects isn't deterministic. For example:

from dask.base import tokenize

for _ in range(3):
    print(tokenize(range(5)))

outputs:

d424666834140c08de99b23f85a7b60c
a3a1bdd9013db6364668f1d224d4cac6
13142f6359eb57df2fc8230942f1ece9

This PR adds a deterministic tokenization for range objects based on their start, stop, and step attributes. With the changes in this PR, the above snippet outputs a consistent hash:

3b69507db0ea8dfcb39cfce774a1df29
3b69507db0ea8dfcb39cfce774a1df29
3b69507db0ea8dfcb39cfce774a1df29

• Tests added / passed
• Passes black dask / flake8 dask

[jrbourbeau]

For reference, the test failures on the AppVeyor are unrelated to the changes in this PR (xref #5948)

[quasiben]

Looks good! I suppose we could do the same for np.arange i the numpy section -- though I would be hard pressed to come up with a use case

[TomAugspurger]

I'm not sure the same can be done for np.arange, since we're tokenizing the output of arange, which is a regular ndarray (IIUC).