[FEA] Adjust libcudf to use kvikIO for small host reads

[GregoryKimball]

Is your feature request related to a problem? Please describe.

For parquet columns that are highly compressible, the page size can be very small. For example, we observe 50-100 KB pages for the NDS-H column l_shipinstruct in lineitems with ZSTD compression. This column has a cardinality of 4, encodes well with run_length, and (may) show other compression opportunities as well.

>>> df['l_shipinstruct'].unique()
0    DELIVER IN PERSON
1     TAKE BACK RETURN
2                 NONE
3          COLLECT COD

When performing IO on these pages, libcudf falls back to a host read from pageable instead of a pinned read via kvikIO.

For an async memory resource, using host read from pageable is not that different than the kvikIO option. However, when using a managed memory pool resource, host read from pageable appears to be ~2x slower in the IO step.

Describe the solution you'd like
We should consider adjusting libcudf to always use kvikIO instead.

I believe we should consider refactoring datasource to not use a read threshold, and also decouple the gds labels from dispatch to kvikIO for host reads.

static constexpr size_t _gds_read_preferred_threshold = 128 << 10;  // 128KB

Additional context

We've collected evidence that with a managed memory pool, the HtoD host copy from pageable to managed is slower than the copy from pageable to device, at least for these small pages.

[GregoryKimball]

@brandon-b-miller @vuule thank you for the discussion today about the pageable copy difference with a managed pool MR.

[vuule]

Related issue #17228: use kvikIO for host reads.
Most likely not required to avoid the pageable copies in this case.

[vuule]

Opened #17260 with a potential fix.
Haven't evaluated the performance impact.