fix(cypher): support connected-prefix carried-scalar reentry (#1004) - #1005

Summary

Support the narrow direct-Cypher MATCH ... WITH ... MATCH ... continuation where the WITH carries one whole-row alias plus scalar outputs derived from that same alias, even when the prefix MATCH is not a literal single-node seed.

Closes #1004.

Scope

This PR intentionally stays narrow:

• supports connected-prefix carried-scalar reentry such as MATCH (a)-[:R]->(b) WITH b, b.id AS bid MATCH (b)-[:S]->(c) ...
• preserves the existing out-of-scope boundaries for broader multi-alias row lowering
• does not attempt the broader #981 multi-binding projection family
• does not attempt the broader #999 multi-stage row-carrier architecture

Implementation

• remove the over-strict lowering guard that required carried-scalar reentry to come from a literal single-node prefix MATCH seed
• add focused coverage for:
  • connected-prefix carried-scalar reentry
  • multiple carried scalar columns from the same connected prefix
  • cross-alias / multi-alias nearby failfast boundaries
  • cudf execution for the exact new connected-prefix shape

Validation

Local:

• PYTHONPATH=. pytest -q graphistry/tests/compute/gfql/cypher/test_lowering.py -> 456 passed, 46 skipped
• PYTHONPATH=. pytest -q graphistry/tests/compute/test_gfql.py -k "reentry" -> 6 passed, 3 skipped, 41 deselected
• ruff check graphistry/compute/gfql/cypher/lowering.py graphistry/tests/compute/gfql/cypher/test_lowering.py -> pass
• ./bin/typecheck.sh -> pass

DGX / cuDF:

• official RAPIDS 26.02 cuda13 targeted run on dgx-spark
• graphistry/tests/compute/gfql/cypher/test_lowering.py::test_string_cypher_executes_with_match_reentry_carried_scalars_from_connected_prefix_shape_on_cudf -> passed

Benchmark-facing result

Fresh origin/master still hits #973 earlier on official interactive-short-2, so this PR by itself does not change current master benchmark status yet.

But when this change is combined with the closed connected-multihop PR head (6b17cf642), the motivating official query moves past the old single-node prefix MATCH seed blocker and reaches the next boundary:

• Cypher MATCH after WITH currently supports a single trailing MATCH pattern

So this lane is still a real benchmark unblocker; it just needs the connected-multihop lane underneath it.