Record Slurm submit failures in the task stderr file

[stark256-spec]

Closes #685.

Problem

When a task ends with a submit failure in Slurm (sbatch returns non-zero), Slurm never starts the job and therefore never writes the task's .err file in the workflow directory. The submit failure reason is raised as a WorkerError, but there is no persistent record in the workflow directory of what caused the failure, so a user inspecting the run afterwards has nothing to look at.

Fix

On a submit failure, capture the sbatch stderr and append it to the same path the task's stderr would have used (as resolved by resolve_stdout_stderr) — the natural place a user looks for a task's error output. The WorkerError is still raised as before.

• New worker_utils.write_submit_failure(stderr_path, message): appends "Slurm job submission failed:\n<message>" to the task's stderr path. It is best-effort — an OSError while writing is logged rather than raised, so it can never mask the original submit error.
• BaseSlurmWorker.submit_job calls it before raising when sbatch returns a non-zero exit code.

This addresses @pagrubel's note on the issue about giving the failure somewhere to live; writing to the task's .err path means it lands where the user already expects task error output rather than in separate metadata.

Tests

Added to beeflow/tests/test_slurm_worker.py (all runnable without a live Slurm/slurmrestd):

• test_write_submit_failure_creates_record — the reason is written to the stderr path.
• test_write_submit_failure_appends — writing appends rather than truncating existing output.
• test_submit_job_records_failure — a failed sbatch (mocked) makes submit_job raise WorkerError and leaves the reason in the task's resolved .err file.

$ pytest beeflow/tests/test_slurm_worker.py -k "write_submit or records_failure"
3 passed

pylint --rcfile=setup.cfg on the changed files: 10.00/10 (source), and the new test code adds no new warnings.