fix(cypher): support connected comma fanout - #1035

Refs #1000

Stacked on #1034.

This Phase 6a slice turns the earlier comma-fanout blocker into a runtime fix. Connected non-linear comma MATCH shapes now lower through joined bindings rows instead of the old single-linear-path normalizer.

What's included:

• connected non-linear comma-pattern join lowering for star-style fanout shapes
• scalar-prefix reentry support through that joined fanout lane
• exact IC6 compile progression now passes the earlier comma-pattern blocker locally
• test amplification for reentry fanout execution, grouped count(post), direct non-reentry fanout execution, and cuDF parity coverage
• CHANGELOG.md prep

What remains adjacent but separate:

• linear connected multi-pattern grouped aggregate overlap still remains fail-fast (group by one alias, aggregate another on already-linear connected fanout shapes)
• that should be the next stacked slice instead of being conflated with this fanout fix

Focused validation run locally on April 2, 2026:

• python3.12 -B -m pytest -q graphistry/tests/compute/gfql/cypher/test_lowering.py -k "connected_star_comma_fanout or issue_1000_ic6_compiles_after_phase6_connected_star_fanout or connected_multi_pattern_grouped_aggregate_overlap or scalar_only_prefix_with_match_reentry" -> 12 passed, 2 skipped, 632 deselected
• python3.12 -B -m pytest -q graphistry/tests/compute/gfql/cypher/test_lowering.py -k "connected_multi_pattern or connected_star_comma_fanout or scalar_only_prefix_with_match_reentry or multiple_post_with_where_clauses or post_with_match_collect_unwind_match or issue_1000_ic6_compiles_after_phase6_connected_star_fanout" -> 30 passed, 4 skipped, 612 deselected
• python3.12 -B -m pytest -q graphistry/tests/compute/gfql/cypher/test_parser.py -k "reentry or unwind" -> 11 passed, 84 deselected
• ruff check graphistry/compute/gfql/cypher/lowering.py graphistry/compute/gfql_unified.py graphistry/tests/compute/gfql/cypher/test_lowering.py graphistry/tests/compute/gfql/cypher/test_parser.py -> passed

Targeted dgx-spark note:

• attempted a narrow RAPIDS/cuDF check for the new star-fanout lane on April 2, 2026
• the remote container failed immediately with cudaErrorMemoryAllocation / std::bad_alloc while constructing a tiny cuDF frame, so that result was operationally inconclusive rather than a clear product failure