Enabled Weighted Sampling

[mkolodner-sc]

Summary

Adds native weighted edge sampling to GiGL's distributed training pipeline via GLT's CPUWeightedSampler. When enabled, neighbors are sampled proportionally to edge weights rather than uniformly.

New API

• DistPartitioner.register_edge_weights(edge_weights) — registers a 1D per-edge weight tensor (homogeneous or dict[EdgeType, Tensor] for heterogeneous) before calling partition_edge_index_and_edge_features(). Weights are partitioned alongside edge features in the same pass (co-partitioned, mirroring the node features + labels pattern).
• build_dataset(weight_edge_feat_name=...) — accepts the name of an existing edge feature column to use as weights. The column is sliced out and removed from the feature tensor before registration, so it is never duplicated in memory.
• DistNeighborLoader(with_weight=True) / DistABLPLoader(with_weight=True) — enables weighted sampling. Defaults to False; must be set explicitly.
• BaseDistLoader.validate_with_weight() — shared validation: raises ValueError if with_weight=True but no weights are registered in the dataset; raises NotImplementedError if used with PPRSamplerOptions (weight-proportional PPR residual propagation is deferred to a future PR).

Implementation notes

• GraphPartitionData.weights (field already existed) carries the partitioned weight tensor to DistDataset._initialize_graph(), which forwards it to GLT's init_graph(edge_weights=...).
• DistDataset.has_edge_weights property reflects whether weights were registered at construction time.
• SamplingConfig.with_weight is now threaded through from the loader rather than hardcoded to False.

Tests

New test file tests/unit/distributed/distributed_weighted_sampling_test.py with 8 tests:

• Correctness (homogeneous + heterogeneous): weight=0 edges to "bad" nodes are never traversed in sampled subgraphs — verified by encoding node class in features and asserting no bad node appears after weighted sampling.
• Partitioner edge cases: features only, weights only, neither, both (with consistency check that GraphPartitionData.edge_ids == FeaturePartitionData.ids), and heterogeneous partial weights (one edge type weighted, another not).