Benchmark the temporal type conversion changes of 4.2.0 #1287

Moved from #1221 (part of #1212), whose documentation part was done in #1284.

Summary

Measure the elapsed time and peak RSS of the readers whose temporal conversion changed in 4.2.0, on the commit planned for 4.2.0, in the combined benchmark #1288 with #1159 and #1162.

This run measures 4.2.0 only. The comparison with the code before the changes comes from the local measurements of each PR (below). The combined run adds absolute values against real S3 and at the benchmark's scales.

Scope

Readers and paths whose conversion changed

• PandasCursor CSV results, whole and chunked, with the C, python, and pyarrow engines: timestamp, date, and timestamp with time zone are converted per column after reading (Give PandasCursor's temporal columns one dtype per Athena type #1273). The fetch methods also convert the temporal columns to datetime and date (Give PandasCursor's temporal columns one dtype per Athena type #1273).
• ArrowCursor CSV results: date is read as date32, and the NULL time, time with time zone, and timestamp with time zone text is replaced with null (Read CSV dates as date32 and temporal text NULL as null in ArrowCursor #1236).
• PolarsCursor CSV results, chunked: the timestamp text is converted in the scan (Convert the timestamp text of chunked Polars CSV reads in the scan #1234).
• UNLOAD results of PandasCursor, ArrowCursor, and PolarsCursor: INT96 timestamps are read at microseconds (Read UNLOAD timestamps at microseconds #1233, Remove the Polars 1.x pyarrow fallback for whole UNLOAD reads with timestamps (GH-1280) #1282). This includes the Polars 2.0 Map read as a list of structs and cast back.
• The fetch methods of Cursor, DictCursor, S3FSCursor, ArrowCursor, and PolarsCursor: the column-naming DataError (Name the column in DataError from the fetch methods of every cursor #1277) and the time zone checks (Raise DataError for time zone values without one instant #1224).

Harness change

The current shapes (flat, nested, json) keep "timestamp" as a bigint (benchmarks/pyathena_bench/config.py, benchmarks/DATASET.md), so none of them converts a temporal value. Add a temporal shape:

• Columns, in config.py: derived from the same snapshot, so that no new input data is needed. For example:
  • a timestamp(3) from from_unixtime("timestamp" / 1e6);
  • a date;
  • timestamp with time zone columns in UTC, in a zone name, and in an offset;
  • time, if useful;
  • the flat scalar columns, so the result keeps a realistic width.
• Cases, in cases.py: add the shape to SHAPES and matrix(). UNLOAD cannot write time or with time zone types, or timestamps other than TIMESTAMP(3). Either give UNLOAD cases a projection of TIMESTAMP(3) and date only, or mark them unsupported, as the json shape does.
• Update the harness tests under benchmarks/tests, benchmarks/README.md, and benchmarks/DATASET.md.
• Merge this before the commit planned for 4.2.0, because the run measures that commit.

Out of scope: running the harness against v4.1.0. The harness uses the workspace's PyAthena, and the PR measurements already cover the comparison.

Local measurements so far (before and after)

The PRs measured on local files, without S3:

• Read UNLOAD timestamps at microseconds #1233: 5M rows, INT96 timestamps and strings. Polars 2.0, ArrowCursor, and PandasCursor whole reads did not change beyond run-to-run variation. The Polars 1.x pyarrow read went from 17-19 ms / 339-354 MiB to 60-78 ms / 593-671 MiB; Remove the Polars 1.x pyarrow fallback for whole UNLOAD reads with timestamps (GH-1280) #1282 removed that read.
• Convert the timestamp text of chunked Polars CSV reads in the scan #1234: chunked Polars CSV reads became faster, and their peak RSS grew with the text (table in the PR).
• Read CSV dates as date32 and temporal text NULL as null in ArrowCursor #1236: ArrowCursor CSV, 5M rows: 0.38-0.54 s → 0.60-0.64 s; the peak RSS ranges overlap.
• Give PandasCursor's temporal columns one dtype per Athena type #1273 (1M rows): PandasCursor CSV reads became 15-79% faster, with peak RSS from -10% to +12%, except a result of only a timestamp with time zone column (+94% to +104%, the text held until conversion). UNLOAD reads went from 0.05-0.06 s to 0.10-0.12 s, about 0.05 s of which is the cast of the date column. fetchall() became 68% faster with 19% less peak RSS.
• Name the column in DataError from the fetch methods of every cursor #1277: +7% for GetQueryResults rows, +8% for DictCursor rows, +2-3% for S3FS rows, ±0 for Arrow.
• Raise DataError for time zone values without one instant #1224: per-value cost of the time zone checks in _to_datetime_with_tz(), +3% for offsets and +10-11% for UTC and zone names.