fix: add retry with exponential backoff to download_file_with_lock

[icenfly]

Fixes #554

Impact

Prevents transient network failures (DNS resolution, connection timeout, etc.) from crashing multi-GPU training runs. Without this fix, a single network hiccup during eval_bundle.zip download at step 2000 causes all 8 ranks to hang and abort via NCCL watchdog, losing hours of training progress.

Change

Added retry logic with exponential backoff and explicit timeout to download_file_with_lock in nanochat/common.py, matching the existing pattern used by download_single_file in nanochat/dataset.py:
• max_attempts = 5 with exponential backoff (2^attempt seconds)
• timeout=30 on urlopen (same as dataset.py's requests.get..., timeout=30))
• Partial file cleanup on failure
• Consistent error message format: Attempt {attempt}/{max_attempts} failed for {filename}: {e}

On a healthy network, the download succeeds on the first attempt and behavior is identical to the original. The existing FileLock mechanism for multi-rank co ordination is unchanged.