Avoid NumPy scalar string representation in tokenize

[jrbourbeau]

This PR updates how tokenize operates on NumPy scalars to avoid using array string representations.

Since NumPy allows users to customize how arrays are printed, the current tokenize behavior can lead to non-deterministic hashes for NumPy scalars in some edge cases. Instead, here we use x.item() to convert NumPy scalars to Python scalars which are then used in tokenize. This allows us to cover edge cases where str(x) has been modified without a performance degradation. With the changes in this PR we have:

In [1]: import numpy as np

In [2]: import dask

In [3]: %timeit dask.base.tokenize(np.array(1.23))
13 µs ± 273 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

and on the current master branch:

In [3]: %timeit dask.base.tokenize(np.array(1.23))
16.2 µs ± 372 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

cc @jcrist if you get a moment to take a look at this or have any thoughts on the topic

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

[jcrist]

This makes sense to me. Didn't know numpy supported redefining printers.

[jrbourbeau]

Thanks for reviewing @jcrist @TomAugspurger