## Summary

This PR adds a `metadata.jsonl` file to enable the HuggingFace Dataset Viewer
to display all 35,586 specimens instead of just ~2,700.

### Problem

The current `imagefolder` auto-detection uses high-resolution original images
(~1.9 MB each), causing the 5GB Parquet conversion limit to be reached after
only 2,700 rows (7.5% of the dataset).

### Solution

The `metadata.jsonl` file explicitly points to thumbnail images instead of
originals. Thumbnails are much smaller, allowing the full dataset to be
previewed in the Dataset Viewer while keeping original high-resolution
images available for download.

### Testing

Tested on a subset at [gwtaylor/massid45-specimens](https://huggingface.co/datasets/gwtaylor/massid45-specimens)
— the Dataset Viewer correctly shows all 20 test rows with thumbnail images.
I didn't want to download and re-upload 66GB so that is why I used the subset test!

### Changes

- **Added:** `metadata.jsonl` — 35,586 entries mapping specimens to thumbnail paths

### Notes

- Original images remain accessible at `images/originals/[fieldid]/[processid].jpg`
- Thumbnails are used only for the Dataset Viewer preview
- No existing files are modified or removed