Add sparse edge attribute bias to graph transformer

[yliu2-sc]

Summary

Add an opt-in sparse edge-attribute attention-bias path for Graph Transformer.

This introduces edge_attr_attention_bias_mode="sparse_linear" on GraphTransformerEncoder, while preserving the default "none" behavior. When enabled, GT uses sampled edge features as sparse per-head additive attention-logit bias.

Changes

• Add edge_attr_attention_bias_mode to GraphTransformerEncoder.
• Support "sparse_linear" mode using positive dims from edge_type_to_feat_dim_map.
• Add zero-initialized Linear(edge_feat_dim -> num_heads, bias=True) per positive-dim edge type.
  • Weight learns edge-feature effects.
  • Bias learns pure edge-presence effects.
• Extend heterodata_to_graph_transformer_input(...) to return sparse edge-attr payloads:
  • pairwise_edge_attr_indices: (batch_idx, query_pos, key_pos)
  • pairwise_edge_attr_values: raw matched edge_attr rows
• Keep edge attrs sparse until final index_put_(..., accumulate=True) into attention bias.
• Use stored edge direction: src -> dst maps to key=src, query=dst.
• Skip zero-dim edge types before matching.
• Accumulate duplicate edge contributions.

Tests

• Added transform tests for sparse edge-attr extraction, stored-direction orientation, padding exclusion, zero-dim skipping, and malformed/missing edge attrs.
• Added encoder tests for zero-init baseline parity, learned presence/feature bias, duplicate accumulation, zero-dim skipping, and PPR compatibility.

Validation

• ruff check passed on touched files.
• git diff --check passed.
• Ad hoc sparse edge-attr validation passed.

Where is the documentation for this feature?: N/A

Did you add automated tests or write a test plan?

Updated Changelog.md? NO

Ready for code review?: NO