Create temp_dir if missing and fail all ranks together instead of deadlocking

[Sophia Wen (hfwen0502)]

Summary

_make_sbd_dir passed temp_dir straight to tempfile.mkdtemp, which does not create parent directories. Passing a temp_dir (the --temp_dir option) that points at a path which does not exist yet therefore raised FileNotFoundError on rank 0.

Because that failure happens before the mpi_comm.bcast that hands the created directory to the other ranks, ranks 1..N block in the broadcast forever. Under MPI the whole job hangs with no output instead of erroring; a single-rank run just surfaces the traceback.

Fix

• Create temp_dir and any missing parents on rank 0 (Path(temp_dir).mkdir(parents=True, exist_ok=True)) before mkdtemp, so a caller may point --temp_dir at a fresh scratch location.
• Wrap rank 0's setup and broadcast its outcome as (ok, payload). On failure every rank raises RuntimeError together rather than deadlocking in the broadcast.

No new imports (Path is already imported); no behavior change when temp_dir already exists or is omitted (the omitted case still routes through tempfile.gettempdir()).

Verification

Reproduced and verified with the real function under mpirun -np 2:

• Before: non-existent --temp_dir → rank 0 raises, rank 1 blocks in bcast; job killed by mpirun --timeout at 8s.
• After, missing dir: both ranks return the same, existing directory; exit 0 in ~1.3s.
• After, uncreatable dir (parent is a regular file): both ranks raise RuntimeError together with a clear message; exit fast, no deadlock.