Differentiate non-missing GT pairwise bias

[yliu2-sc]

Summary

This PR differentiates non-missing pairwise GT attention-bias entries from missing entries.

Previously, sparse pairwise attention-bias values collapsed explicit zero-valued entries, diagonal self-pairs, and missing entries into the same 0.0 value. This makes it hard for the GT attention mechanism to distinguish self relationships from absent pairwise structure.

Changes

• Adds pairwise_nonmissing_indices to Graph Transformer sequence auxiliary data.
• Treats valid diagonal self-pairs as non-missing, even when their pairwise value is 0.0.
• Treats explicit sparse CSR entries as non-missing, including explicit zero-valued entries.
• Leaves missing non-self pairs as 0.0 without non-missing bias.
• Adds a zero-initialized learned per-head non-missing attention bias in GraphTransformerEncoder.
• Requires multiple pairwise attention-bias attributes to share identical non-missing support.

Tests

• Added transform tests for:
  • diagonal self-pairs marked non-missing
  • explicit sparse entries marked non-missing
  • missing non-self pairs excluded
  • padding excluded
  • mismatched pairwise attr support raising ValueError
• Added encoder test for:
  • head-specific non-missing bias
  • only indexed pairs receiving the bias
  • missing pairs remaining unchanged

Verification

Passed:

• Custom smoke test for transform support coordinates and encoder non-missing bias
• ruff check on touched files
• ruff format --check on touched files
• git diff --check

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