[REVIEW] Add Series.diff() via Numba kernel

[beckernick]

This PR adds the diff functionality to numeric cudf.Series, analogous to pandas.

Summary of Changes

• Adds a diff method to cudf.Series, matching pandas API
• Adds a gpu_diff numba kernel in cudautils for out-of-place differencing
• Tests for forward and backward differencing of numeric dtypes
• Cell values that would become NaN based on the period in pandas diff are currently set to be -1 in cudf diff
• Requires that a column not contain any null values

This could serve as a stopgap until diff can be implemented in libcudf (#1271 ), as the performance is similar to standard binary operations on columns of millions of rows.

import cudf
import numpy as np

nelem = int(1e7)
df = cudf.DataFrame({'a':np.random.sample(nelem)})

%timeit df.a.diff(2)
%timeit df.a + 1
4.33 ms ± 143 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
5.47 ms ± 40.2 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)