Use pinned bounce buffer in the CSV reader

[vuule]

Description

Decompressed input data in the CSV reader is copied to the device in chunks.
This PR modifies the process to copy each chunk to a pinned bounce buffer, from which the data is then copied to the device. This should lead to higher throughput of the H2D copies.

Checklist

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

[vuule]

No perf numbers because my dev system has oddly low throughput that does not change with pinned memory.
Profiles from a lab system show big improvement.

[GregoryKimball]

Thank you @vuule for preparing this option for switching read_csv to use a pinned buffer for mmap file reads. Here are my quick benchmarking results read files containing LLM-inspired data. I compared recent 23.08 commit d4c2d1ccee0 with fc7b81690b18 on this branch. Solid points use the pageable mmap baseline and open circles use the pinned mmap proposal.

The CSV format is the only one expected to change from this diff. We see perhaps a 15% runtime increase from the largest file sizes. I will update this issue after we discuss the merits internally.

[vuule]

Closing based on results in the comment above