Fix checkpoint directory race condition in DDP training

[0xadvait]

Problem

Multi-GPU training with torchrun and --overwrite crashes intermittently: every rank executes the checkpoint-directory setup in train_loop, so all ranks call shutil.rmtree on the same directory at once, and ranks that lose the race die with FileNotFoundError (full traceback in #868).

Fix

• Only the main process deletes (--overwrite) and creates the checkpoint directory.
• All ranks synchronize on dist.barrier() after directory setup, so no rank proceeds before the directory state is settled.
• Resume detection stays on all ranks (read-only), and the resume error branches still raise on every rank, so non-main ranks cannot continue into a broken run.

Verification

• ruff check and ruff format --check pass.
• Simulated the race standalone with 8 processes against a shared directory: the unguarded pattern hit FileNotFoundError 140 times across 20 attempts, the guarded pattern produced 0 errors across 50 attempts.
• save_checkpoint already guards on is_main, so no other rank-unsafe filesystem operations remain in the script.

Fixes #868