perf(arrow-cast): improve custom temporal formatting performance in ArrayFormatter

[lyne7-sc]

Which issue does this PR close?

• Closes #NNN.

Rationale for this change

ArrayFormatter currently reparses custom Chrono format strings for every temporal value.

As an ArrayFormatter is typically created once and reused for an array, the parsed format items can instead be prepared once when constructing the formatter and reused for every value.

What changes are included in this PR?

• Precompile custom temporal format strings when constructing an ArrayFormatter
• Reuse the compiled format items for date, time, and timestamp values
• Add benchmarks covering default and custom temporal formats

Are these changes tested?

Yes.

Are there any user-facing changes?

No. There are no public API or behavior changes.

Benchmarks

group                                       main                                    optimized
-----                                       ----                                    ---------
format_temporal/date64/custom_long          1.79  1559.4±47.26µs        ? ?/sec     1.00   872.7±32.05µs        ? ?/sec
format_temporal/date64/custom_short         1.68   797.6±39.91µs        ? ?/sec     1.00   473.7±15.21µs        ? ?/sec
format_temporal/date64/default              1.02   464.7±12.82µs        ? ?/sec     1.00    454.9±8.67µs        ? ?/sec
format_temporal/timestamp/custom_long       1.77  1635.8±175.75µs       ? ?/sec     1.00   925.6±29.74µs        ? ?/sec
format_temporal/timestamp_tz/custom_long    1.54  1939.6±42.35µs        ? ?/sec     1.00  1259.3±41.78µs        ? ?/sec

[Jefffrey]

can reproduce results locally 👍

format_temporal/date64/default
                        time:   [321.48 µs 322.23 µs 323.09 µs]
+                       change: [−2.6583% −2.0392% −1.4818%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 11 outliers among 100 measurements (11.00%)
  7 (7.00%) high mild
  4 (4.00%) high severe
format_temporal/date64/custom_short
                        time:   [515.48 µs 517.51 µs 519.66 µs]
+                       change: [−21.775% −21.467% −21.154%] (p = 0.00 < 0.05)
                        Performance has improved.
format_temporal/date64/custom_long
                        time:   [880.49 µs 882.41 µs 884.21 µs]
+                       change: [−32.212% −31.955% −31.705%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 1 outliers among 100 measurements (1.00%)
  1 (1.00%) high mild
format_temporal/timestamp/custom_long
                        time:   [934.32 µs 937.06 µs 939.70 µs]
+                       change: [−29.966% −29.722% −29.440%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 3 outliers among 100 measurements (3.00%)
  3 (3.00%) high mild
Benchmarking format_temporal/timestamp_tz/custom_long: Warming up for 3.0000 s
Warning: Unable to complete 100 samples in 5.0s. You may wish to increase target time to 6.0s, enable flat sampling, or reduce sample count to 60.
format_temporal/timestamp_tz/custom_long
                        time:   [1.2028 ms 1.2077 ms 1.2126 ms]
+                       change: [−24.125% −23.734% −23.338%] (p = 0.00 < 0.05)
                        Performance has improved.

[lyne7-sc]

@Jefffrey thanks for your review!