Add NuBench training datasets

[sevmag]

Closes #887.

Summary

• Adds NuBenchDataset (in src/graphnet/datasets/nubench_datasets.py), a single ERDAHostedDataset entry point for the NuBench benchmark suite (arXiv:2511.13111).
• Ships a registry of the available NuBench datasets — cluster, flower_l, flower_s, flower_xl, hexagon, hexagon_ice_le — each pinned to its ERDA hash and corresponding NuBenchDetector subclass.
• Handles the NuBench split convention automatically: train/val read from the merged_photons pulsemap, test reads from pulses_no_noise.
• Re-exports NuBenchDataset, NuBenchSpec, FEATURES_NUBENCH, and TRUTH_NUBENCH from graphnet.datasets.

The NuBench detector classes (Cluster, FlowerL, FlowerS, FlowerXL, Hexagon, Triangle) and geometry tables already exist on main; this PR adds the missing Dataset layer so the benchmarks can actually be trained on.

Usage

from graphnet.models.graphs import KNNGraph
from graphnet.models.detector.nubench import Hexagon
from graphnet.datasets import NuBenchDataset

ds = NuBenchDataset(
    name="hexagon_ice_le",
    download_dir="/path/to/nubench_data",
    data_representation=KNNGraph(detector=Hexagon()),
)

Test plan

• CI passes (lint + unit tests).
• Smoke-test NuBenchDataset(name="hexagon_ice_le", ...) end-to-end: download from ERDA, build train/val/test splits, iterate a few batches.
• Verify available_datasets() lists all entries in the registry.

[Aske-Rosted]

I have a single comment, otherwise this seems fine to me.