Derive with_edge from edge feature presence (perf)

[kmontemayor2-sc]

What

BaseDistLoader.create_sampling_config hardcoded with_edge=True for every
loader. This derives it from whether the dataset actually has edge features:

  with_edge = dataset_schema.edge_feature_info is not None

DatasetSchema.edge_feature_info is already populated in both colocated and
graph-store modes (fetched from storage rank 0 in graph-store mode), so the
derivation is correct on compute nodes as well.

Why

For datasets with no edge features anywhere, with_edge=True forces GLT to
collect sampled edge ids per hop and transport them in every SampleMessage.
Those eids are pure overhead when there are no edge features to gather: the
sampler writes an int64 eid tensor of one element per sampled edge (gated on
self.with_edge). With with_edge=False, GLT takes the sample path instead
of sample_with_edge, output.edge is None, no eids are collected or
serialized, and edge-feature-store construction is skipped.

Behavior change (changelog)

Batches produced from featureless datasets no longer carry GLT's data.edge
(homogeneous) or data[etype].edge (heterogeneous) sampled-edge-id attribute.
A repo-wide grep found no consumer that reads this attribute off a produced
batch — the only in-repo references are the sampler's own write sites. External
code that reads batch.edge on a featureless dataset would now find it
absent/None. Datasets that do have edge features are unaffected.

Tests

• Direct unit tests for create_sampling_config: with_edge is False when
  edge_feature_info is None, True for homogeneous and heterogeneous edge
  features.
• End-to-end: a featureless homogeneous dataset yields batches with no sampled
  edge ids (data.edge is None).
• Weighted sampling without edge features still avoids zero-weight edges.
• PPR unaffected (outputs come from metadata, not sampled eids).
• Graph-store provenance: compute-side edge_feature_info is populated from the
  storage cluster.

All affected files pass; make type_check passes.

[kmontemayor2-sc]

/all_test

[github-actions]

GiGL Automation

@ 00:38:38UTC : 🔄 Lint Test started.

@ 00:46:08UTC : ✅ Workflow completed successfully.

[github-actions]

GiGL Automation

@ 00:38:38UTC : 🔄 E2E Test started.

@ 02:13:17UTC : ❌ Workflow failed.
Please check the logs for more details.

[github-actions]

GiGL Automation

@ 00:38:39UTC : 🔄 C++ Unit Test started.

@ 00:40:22UTC : ✅ Workflow completed successfully.

[github-actions]

GiGL Automation

@ 00:38:41UTC : 🔄 Python Unit Test started.

@ 01:54:03UTC : ✅ Workflow completed successfully.

[github-actions]

GiGL Automation

@ 00:38:42UTC : 🔄 Scala Unit Test started.

@ 00:48:27UTC : ✅ Workflow completed successfully.

[github-actions]

GiGL Automation

@ 00:38:42UTC : 🔄 Integration Test started.

@ 02:02:27UTC : ✅ Workflow completed successfully.

[mkolodner-sc]

Thanks Kyle for this optimization! Generally LGTM with some minor comments