unc-liver: add the split CSV, center the hero image

Syncs the unc-liver card on the Hub with the reviewed state on GitHub (open-h/OpenH-RF).

The Splits section said `dataset_split.csv` ships alongside the HDF5 files, but it was never uploaded. This adds it as `assets/dataset_split.csv`: one row per HDF5 file (`hdf5_filename`, `volume`, and a `<split>_train` / `<split>_valid` flag pair per split). Checked against the release: all 1906 files in `data/` are listed, the train/valid counts match the card's table, and no volume is in both train and valid for any split.

The card now links the CSV and describes its columns, corrects the fold "neither train nor validation" range from 97-153 to 97-113, and centers the hero image at a fixed 480 px width instead of filling the page.

**GitHub PR:**
- [tristan-deep/OpenH-RF#35](https://github.com/tristan-deep/OpenH-RF/pull/35) — unc-liver: add the split CSV, center the hero image

Files changed: 2.

```
+ unc-liver/assets/dataset_split.csv
~ unc-liver/README.md
```

unc-liver/README.md

@@ -21,7 +21,9 @@ size_categories:
 
 # fullwave-abdominal-wall: Fullwave Abdominal Wall Simulation Dataset
 
-![Scan-converted B-mode through a simulated abdominal wall](assets/main.png)
+<p align="center">
+  <img src="assets/main.png" alt="Scan-converted B-mode through a simulated abdominal wall" width="480">
+</p>
 
 *Scan-converted B-mode reconstructed from the full-synthetic-aperture channel data of one simulated acquisition in [`data/`](https://huggingface.co/datasets/nvidia/OpenH-RF/tree/main/unc-liver/data): the abdominal wall layers in the near field above the liver.*
 
@@ -136,7 +138,7 @@ All maps are on the polar reconstruction grid: 640 radial samples spanning 5.0
 
 ### Splits
 
-`dataset_split.csv` ships alongside the HDF5 files. Splits are **grouped by phantom volume** so that no volume appears in both train and validation. This avoids anatomical leakage, which matters because many acquisitions are different 2-D slices of the same volume.
+[`assets/dataset_split.csv`](assets/dataset_split.csv) assigns every acquisition to the splits below, one row per HDF5 file (`hdf5_filename`, `volume`, and a `<split>_train` / `<split>_valid` flag pair per split). Splits are **grouped by phantom volume** so that no volume appears in both train and validation. This avoids anatomical leakage, which matters because many acquisitions are different 2-D slices of the same volume.
 
 | Split | Train | Validation |
 |---|---|---|
@@ -147,7 +149,7 @@ All maps are on the polar reconstruction grid: 640 radial samples spanning 5.0
 | `fold3` | 1314 | 488 |
 | `fold4` | 1328 | 465 |
 
-Note the `fold*` splits are **not** complementary: between 97 and 153 acquisitions are in neither the train nor the validation set of a given fold.
+Note the `fold*` splits are **not** complementary: between 97 and 113 acquisitions are in neither the train nor the validation set of a given fold.
 
 ## Subject Metadata