Support constructing pylibcudf Columns and Tables from views into arbitrary objects

[vyasr]

Description

The new constructors are necessary when we need to produce views into data that is not owned by one of the standard libcudf types (column/table). The original motivating example is contiguous_split, so that API is modified in this PR for testing.

Closes #17040
Supersedes #17543

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[vyasr]

Since this change removes a copy in contiguous_split, I did a quick benchmark to confirm the expected impact, especially since this API is important for the distributed Polars engine work. Here's my (very simple) benchmark:

import pyarrow as pa
import pytest

import pylibcudf as plc


@pytest.fixture
def packed_table():
    pa_tbl = pa.table({
        "a": range(int(1e8)),
        "b": range(int(1e8)),
        "c": range(int(1e8)),
    })
    plc_tbl = plc.interop.from_arrow(pa_tbl)
    return plc.contiguous_split.pack(plc_tbl)


def bench_unpack(benchmark, packed_table):
    benchmark(plc.contiguous_split.unpack, packed_table)

and here are the results:

# Before
---------------------------------------------- benchmark: 1 tests ----------------------------------------------
Name (time in ms)        Min      Max     Mean   StdDev   Median      IQR  Outliers      OPS  Rounds  Iterations
----------------------------------------------------------------------------------------------------------------
bench_unpack          8.8083  82.7581  53.5293  23.6607  46.2718  37.7695      39;0  18.6814     108           1
----------------------------------------------------------------------------------------------------------------

# After
----------------------------------------------- benchmark: 1 tests ----------------------------------------------
Name (time in us)        Min      Max    Mean  StdDev  Median     IQR  Outliers  OPS (Kops/s)  Rounds  Iterations
-----------------------------------------------------------------------------------------------------------------
bench_unpack          2.0800  25.2870  2.2061  0.5804  2.1600  0.0421  319;2200      453.2812   33326           1
-----------------------------------------------------------------------------------------------------------------

Note that these are actually being reported in different units, ms on top and us below, so it's about 400x faster. Obviously the results are very dependent on the data size, and in this case I picked a large buffer. If I drop it down to a million instead of a hundred million rows, the "After" timings are largely unchanged as expected (since we are just viewing the buffer), while the "Before" timings look like

-------------------------------------------- benchmark: 1 tests --------------------------------------------
Name (time in ms)        Min     Max    Mean  StdDev  Median     IQR  Outliers       OPS  Rounds  Iterations
------------------------------------------------------------------------------------------------------------
bench_unpack          1.0384  5.0173  1.1214  0.2195  1.0943  0.0596     12;28  891.7214     670           1
------------------------------------------------------------------------------------------------------------

We still see a fairly large gap. For the limiting case of a single-row table, we get

------------------------------------------------- benchmark: 1 tests ------------------------------------------------
Name (time in us)         Min       Max     Mean  StdDev   Median     IQR  Outliers  OPS (Kops/s)  Rounds  Iterations
---------------------------------------------------------------------------------------------------------------------
bench_unpack          52.0620  122.4501  55.8294  7.2951  53.4215  1.8415   249;391       17.9117    3040           1
---------------------------------------------------------------------------------------------------------------------

so there's still about 30x overhead that we get to save ourselves from avoiding the allocation steps altogether.

[bdice]

Approving with one question on pointer types.

[vyasr]

/merge