Persist local embedding progress on failure

[clemsgrs]

Summary

Local single-process embedding used to batch all persistence until the entire embedding stage finished. If a later slide failed, earlier successful slides had been computed in memory but nothing had been written yet, so the run lost all embedding progress.

This change makes local embedding durable and resumable:

• persist each completed slide immediately during local single-process embedding
• keep process_list.csv feature and aggregation status aligned with the artifacts that have actually been written
• when Pipeline.run(...) is rerun with resume=True, skip slides whose local embedding artifacts are already complete and marked successful
• leave the multi-GPU pipeline path unchanged, while still making direct local embed_slides(..., output_dir=...) durable on later-slide failure

User impact

If a large local batch fails near the end, the embeddings that finished before the failure stay on disk instead of being discarded. A resumed local pipeline run now reuses those completed artifacts and computes only the remaining slides.