fix: preserve benchmark agent failures

[abrichr]

What changed

• Return terminal error actions when hosted VLM calls fail.
• Return terminal error actions when API, unified-baseline, or WAA demo responses cannot be parsed.
• Preserve explicit model-issued completion as done.

Why

The benchmark action schema already supports error, and the evaluation adapters handle it as a distinct terminal state. Returning done for provider and parse failures made task completion and agent failure indistinguishable.

Validation

• 470 unit tests passed; 5 skipped.
• 77 focused failure and baseline tests passed.
• Ruff and formatting checks passed for every changed file.
• git diff --check passed.
• The optional integration collection requires openadapt-evals, which is not installed in this worktree environment.