MINOR: Use build-once Test workflow

[xborder]

This converts the Test workflow to a build-once/test-many shape so we can compare it against the existing control workflow.

What changed:

• Build Java artifacts once on Linux/JDK 17 and upload the prepared build tree plus locally installed Arrow Maven artifacts.
• Reuse those artifacts for Java tests on Linux JDK 17/21/23, macOS Intel, macOS ARM, and Windows.
• Build C Data native artifacts once on Linux/JDK 17 and reuse them for C Data tests on Linux JDK 17/21/23.
• Add cheap JDK 21/23 compile checks so newer JDK compile coverage is retained.
• Add a manual workflow_dispatch skip_integration input for timing-only comparisons.
• Remove the separate prototype workflow now that the mechanism is in the real Test workflow.

Validation:

• Test run 27494182611 passed with integration.
• Manual Test run 27495432430 passed with integration skipped.
• check-yaml passed.
• RAT passed.
• git diff --check passed.
• pre-commit run --all-files reached ShellCheck but could not continue locally because Docker is not installed in this environment.

Control comparison:

• Older no-change control 27446267586: 29.4m wall-clock, about 473.9 weighted runner minutes.
• Phase 1-3 baseline 27478899804: 25.6m wall-clock, about 436.5 weighted runner minutes.
• Build-once with integration 27494182611: 31.7m wall-clock, about 148.1 weighted runner minutes.
• Build-once without integration 27495432430: 15.4m wall-clock, about 123.8 weighted runner minutes.

The with-integration build-once run was slower in wall-clock only because integration took 31.6m in that run. The matrix itself is much cheaper: excluding integration, weighted runner minutes dropped from about 411.0 to about 123.8 versus the Phase 1-3 baseline.