feat(pipeline): Support resuming training runs

[hummat]

Problem

Currently there's no way to resume an interrupted training run:

• Without --overwrite: Existing output files trigger the skip logic, so the train step is skipped entirely
• With --overwrite: The existing checkpoint directory is deleted, losing all training progress

Motivation

Long training runs (hours) can be interrupted by:

• GPU crashes / OOM
• System reboots
• User mistakes

Losing progress is frustrating and wastes compute.

Proposed Solution

Add a --resume flag (or similar) that:

1. Does not skip the train step when output exists
2. Does not delete existing checkpoints
3. Passes --trainer.load-dir pointing to the existing run's checkpoint directory

Alternatively, detect if a partial training run exists (checkpoints present but training incomplete) and automatically resume.

Expected Behavior

Users should be able to continue training from a checkpoint without losing progress. The underlying trainer (sdf-train) already supports this via:

• --trainer.load-dir PATH — directory containing checkpoints
• --trainer.load-step INT — specific step to resume from

Alternatives Considered

Current workaround: Users must manually invoke sdf-train with --trainer.load-dir to resume. This works but requires knowing the internal command structure.

Tasks

• Add --resume flag to run.sh
• Modify skip logic to allow resume when flag is set
• Pass --trainer.load-dir to sdf-train when resuming
• Optionally auto-detect latest checkpoint step
• Update README with resume documentation
• Add tests for resume behavior