ci: the test, wheels and generated jobs free runner disk; the test build drops debug info - #776

Includes #775's toolchain repin. It is cherry-picked as its own commit: the four dtolnay/rust-toolchain pins move from b725eda5 to 39008827. #775 was ejected from the merge queue when its test job hit this disk problem, and #776 cannot pass zizmor without the repin, so the two land together here. #775 closes once this merges.

The test job runs out of disk on most open PRs. It already frees about 32 GB and builds with line tables, but on 2026-11-15:

• plan, exec, optimizer, plan-codec: Python UDF relations: IR, frame seam, operators and codec feature (P3-T014a, PR-26a) #746 and session, plan, exec, python, connect: Dataset.observe and Observation (P3-T084, PR-51) #756 started with 119 GB free and ended at 84 MB. The runner crashed with "No space left on device" and kept no log.
• On plan, memory, plan-codec, session, exec: nested literals, a declared file schema and the LocalRelation carrier (P3-T012a part 2a, PR-12) #772, ld died with SIGBUS (signal 7) while linking krishiv-server's tests, with 434 MB left.

#768 fixed only diff-coverage. This PR extends the same treatment to the other heavy jobs.

Changes

• New composite action .github/actions/free-disk. It runs the cleanup list the test, coverage and nightly jobs already use, and also removes the preinstalled container images, as ci: diff-coverage builds without debug info, prunes container images and reports its disk use #768 does. It prints df -h / before and after.
• test:
  • uses the action;
  • builds with CARGO_PROFILE_DEV_DEBUG and CARGO_PROFILE_TEST_DEBUG set to false (panic backtraces keep function names but lose line numbers; no test reads DWARF);
  • after the doctests, removes the workspace run's test executables before the four feature-specific runs build their own;
  • prints df and the size of target there and at the end.
• wheels (both legs): uses the action and prints df at the end.
• generated (docs-site.yml): uses the action, builds without debug info and prints df at the end.

diff-coverage is unchanged. Since #768 its runs end at 73–78% of the disk (#751, #772); those two now fail on the 90% added-line coverage gate itself, not on disk.

Checks

• zizmor 1.30.1, offline, on the three changed files: no findings.
• zizmor 1.30.1 online, on all of .github/ with the repin: no findings.
• cargo xtask workflow-lint: OK.

This PR changes ci.yml, so its own test run shows whether the build now fits: the new df/du lines report it.