GH-50170: [CI][Packaging][Linux] Fix cache

[kou]

Rationale for this change

We want to reduce CI time as much as possible for developer experience and avoiding consuming shared resources.

What changes are included in this PR?

• Use apache/infrastructure-actions/stash instead of actions/cache to use larger cache storage
• Enable ccache
• Increase ccache size because 500MB is small for our case
• Fix Docker image cache
• Disable RapidJSON's ccache detection that causes needless ccache call such as ccache ccache cc ...
• Use Ninja instead of Make

Are these changes tested?

Yes.

CI time is reduced to 20min-60min from 60min-120min.

Are there any user-facing changes?

No.

• GitHub Issue: [CI][Packaging][Linux] Reduce GitHub Actions usage #50170

[github-actions]

⚠️ GitHub issue #50170 has been automatically assigned in GitHub to PR creator.

[kou]

Before: https://github.com/kou/arrow/actions/runs/27544568196/usage
After: https://github.com/kou/arrow/actions/runs/27553127454/usage (Some slow jobs didn't have cache yet. Sorry.)

[kou]

This is for debugging why ccache couldn't cache something.

[kou]

Before: https://github.com/apache/arrow/actions/runs/27590222397/usage?pr=50185 (1d 1h 20m 16s)
After1: https://github.com/apache/arrow/actions/runs/27595144017/usage?pr=50185 (17h 41m 5s)
After2: https://github.com/apache/arrow/actions/runs/27647194242/usage?pr=50185 (15h 18m 38s)

[kou]

+1

25h -> 15h

[conbench-apache-arrow]

After merging your PR, Conbench analyzed the 4 benchmarking runs that have been run so far on merge-commit 39770de.

There were 6 benchmark results indicating a performance regression:

• Commit Run on test-mac-arm at 2026-06-17 05:36:21Z

  • tpch (R) with engine=arrow, format=native, language=R, memory_map=False, query_id=TPCH-12, scale_factor=1

• Commit Run on amd64-m5-4xlarge-linux at 2026-06-17 01:49:54Z

  • dataset-serialize (Python) with dataset=nyctaxi_multi_parquet_s3, format=csv, selectivity=1pc

• and 4 more (see the report linked below)

The full Conbench report has more details. It also includes information about 6 possible false positives for unstable benchmarks that are known to sometimes produce them.