fix: replace removed SampleBaseDataset in kg_emb

[capccode]

Summary

Closes #952

SampleBaseDataset was deleted in the PyHealth 2.0 dataset refactor (commit 62df8d8), breaking all imports of pyhealth.medcode.pretrained_embeddings. Anyone importing the kg_emb module gets an ImportError.

• SampleKGDataset now inherits from torch.utils.data.Dataset directly, inlining the three fields (samples, dataset_name, task_name) and __len__ that the deleted parent provided
• Removed broken from pyhealth.datasets import SampleBaseDataset import and type annotations from 5 model files and the splitter
• Added pandarallel to graph optional dependencies (required by umls.py)

Files changed (8):

• kg_emb/datasets/sample_kg_dataset.py — new parent class + inlined fields
• kg_emb/datasets/splitter.py — removed broken import
• kg_emb/models/kg_base.py, complex.py, distmult.py, rotate.py, transe.py — removed broken import
• pyproject.toml — added pandarallel to [project.optional-dependencies] graph

Test plan

• from pyhealth.medcode.pretrained_embeddings import * no longer crashes
• All 5 models (TransE, RotatE, DistMult, ComplEx, KGEBaseModel) import and instantiate
• SampleKGDataset works with torch.utils.data.DataLoader
• kg_emb.datasets.splitter.split() works
• lm_emb is unaffected (no SampleBaseDataset references)

[jhnwu3]

Ah I see, there's an entirely separate SampleKGDataset here. Can you work with @joshuasteier later after our deadliens here to revamp it towards the graph/ module here.

The problem is we still want it to all follow through the PyHealth pipeline here, i.e like we do with GraphCare's implementation as technically Graphs are not disjoint from any other modality here. We just have a bunch of old technical debt that's stuck here.

Will hopefully get to your GRASP PR at some point.