fix(reproduce): preserve user files when cleaning download caches

[reacher-z]

What does this PR do?

clawbench-reproduce --work-dir <existing-directory> previously deleted the entire supplied directory after comparison, including unrelated user files. Downloads now go into a unique invocation-owned child directory. Only that child is removed when the invocation finishes or unwinds; --keep-cache retains it and prints its exact path.

Temporary-directory cleanup covers PASS, FAIL, download errors, judge errors and interruption. Independent child directories prevent one invocation from deleting another invocation's cache. The caller's parent directory and pre-existing contents are preserved. Includes the required Unreleased changelog entry.

Corpus

• v2
• v1
• both
• not applicable

Host-side reproduction CLI cleanup only.

Test plan

• 11 new offline regressions: five completion/failure modes with and without --keep-cache, plus overlapping cache lifetimes. Each CLI case preserves an unrelated sentinel and a previous cache.
• uv run --frozen pytest -q — 294 passed, 3 skipped.
• Ruff check/format on changed Python files; uv run --frozen pyright — 0 errors.
• git diff --check.
• uv build — wheel and sdist successfully built.

Download and judge calls are mocked in the regression tests; no live HF/model run or package publication was performed. Reproduction metric validation and judge-cache correctness remain separate issues #362–#364. Retained downloads now live one unique directory below --work-dir, so callers inspecting the old direct layout should use the printed cache path.

Related issues

Fixes #361. Part of the next-release quality checklist #366.

@Perry2004 please review the cache ownership/cleanup behavior and prioritize this data-loss fix for the next release.