Migrate string slice APIs to pylibcudf

[brandon-b-miller]

This PR introduces pylibcudf string slice APIs and migrates the cuDF cython to use them. Part of #15162

[charlesbluca]

Thanks @brandon-b-miller! Some small drive-by comments for now:

[charlesbluca]

Not blocking this PR, but wonder if it makes sense to align the libcudf/pylibcudf structure on either substring or slice rather than using both interchangeably (though don't have much context with strings submodules to know if this was an intentional choice)

[brandon-b-miller]

There's a failing test here that could be coming from libcudf - I get different answers for a string with just one space in it depending on if I use the column wise or scalar APIs:

import pyarrow as pa
import cudf._lib.pylibcudf as plc

col = plc.interop.from_arrow(
    pa.array(
        [
            " "
        ]
    )
)

slr_start = plc.interop.from_arrow(pa.scalar(-1))
slr_stop = plc.interop.from_arrow(pa.scalar(-1))

res = plc.strings.slice.slice_strings(col, slr_start, slr_stop)
print(plc.interop.to_arrow(res))

col_start = plc.interop.from_arrow(
    pa.array(
        [
            -1
        ]
    )
)

col_stop = plc.interop.from_arrow(
    pa.array(
        [
            -1
        ]
    )
)

res = plc.strings.slice.slice_strings(col, col_start, col_stop)

print(plc.interop.to_arrow(res))

This prints

[
  [
    ""
  ]
]
[
  [
    " "
  ]
]

Should we expect do get a different result for slice(" ", start=-1, stop=-1) depending on if the -1's come from a scalar or column?

[brandon-b-miller]

Should we expect do get a different result for slice(" ", start=-1, stop=-1) depending on if the -1's come from a scalar or column?

cc maybe @davidwendt for the above

[davidwendt]

The scalar version does operate different from the columnar version when negative values are involved.
Documentation here does not describe this: https://docs.rapids.ai/api/libcudf/stable/group__strings__slice#gaa6f853a3b807391dadad69ad8727bac0

Negative values for the scalar version do a wrap around to be consistent with the pandas string slicing.
https://pandas.pydata.org/docs/reference/api/pandas.Series.str.slice.html
The columnar version does not support negative values in this way. You can pass negative values for the stops column to indicate to-the-end-of-the-string. Otherwise, out-of-bounds position values will result in an empty string.

[vyasr]

Should we update the docs in this PR? That seems like the main action item if the test behavior is correct.

[brandon-b-miller]

Thanks @davidwendt for the context I managed to handle the edge case in 78fe267. All tests should pass here now.

[wence-]

Some small queries

[wence-]

Thanks Brandon

[brandon-b-miller]

/merge