[tdene]

What does this PR do?

Forces Jest invocations in nv-internal eval scripts to run in-band: each jest line in an instance's run_script.sh gains --runInBand (skipped when the row already sets --maxWorkers) and --forceExit.

This has already been done for part of the dataset manually (90 rows), but the rest of the dataset still errors out due to Jest issues, so a comprehensive fix is useful.

Jest sizes its default worker pool from os.cpus(). The eval container gets a memory cgroup but no cpuset, so a single episode on a large node spawns a core-count fleet of Node workers.
The episode runs under rootless Apptainer with a squashfuse rootfs: every node_modules read funnels through one user-space FUSE daemon per container, inside a fixed memory budget, on nodes running many concurrent episodes. Under load the worker fleet starves the FUSE daemon (D-state pileup) or exhausts the cgroup; separately, suites that leak open handles keep the jest process alive after all tests have passed.

Thus we reach the full swebench_tests_timeout, produce no output, and score 0, corrupting reward.

Checklist

• I have read the contributing guidelines.
• The change is focused; unrelated "drive-by" edits are tracked as separate issues/PRs.
• Tests added or updated and pass locally, or N/A for docs-only / non-code changes (so CI unit/server checks pass when applicable).
• Pre-commit checks pass locally (pre-commit run --all-files) (so CI lint/format/copyright pass).
• All commits have DCO sign-off (git commit -s) (so the DCO check passes).

[tdene]

/ok to test a910318