Add ppr_relation_features token input to the PPR Graph Transformer path

[mkolodner-sc]

What

Adds a reserved ppr_relation_features anchor-relative token input to the PPR
Graph Transformer path, and consolidates the reserved-feature validation shared
by the transform and the encoder.

Details

• transforms/graph_transformer.py
  • New PPR_RELATION_FEATURES_NAME token input: the columns of the PPR
    edge_attr beyond the scalar weight (col 0) are gathered per (anchor,
    neighbor) into a (batch, max_seq_len, k) feature tensor in
    _build_sequence_layout_from_ppr_edges and exposed through the anchor
    feature composers.
  • Neighbors stay ordered by descending PPR weight; relation-feature rows are
    reordered by the same permutation so they remain aligned with their edges.
  • Relaxes _validate_ppr_sequence_input so the batch may also carry the
    original relation edge types (which feed the relation-message channel); only
    the "ppr" edges drive the sequence.
• New _validate_reserved_anchor_feature_usage helper centralizes the
  placement/method checks for ppr_weight and ppr_relation_features, replacing
  the duplicated blocks in heterodata_to_graph_transformer_input and
  GraphTransformerEncoder.

Testing

• ruff check / ruff format clean; module compiles.
• Unit-level: validator rejects misuse (pairwise/scalar bias, non-ppr method);
  PPR sequences are weight-ordered with relation features aligned to their edges;
  truncation to max_seq_len keeps the top-weight neighbors.