perf(gfql): generalize seeded typed-hop fast path to polars/polars-gpu (#1755)

[lmeyerov]

Stacked on #1759 (base = perf/gfql-seeded-typed-hop-fastpath-1755). Review/merge #1759 first.

What

Generalizes the #1755 seeded typed-hop cypher fast path (pandas-only in #1759)
to polars and polars-gpu via _seeded_typed_return_dst_polars — the same
seed-first reduction (seed out-edges → typed-edge filter → destination nodes)
expressed with polars filters. It hooks upstream at
_execute_seeded_typed_hop_fast_path, before the lazy pipeline, so none of the
woven binding_rows_polars / hop_polars / combine machinery changes.

dgx measurement (26.02-gfql-polars, 50k Person / 200k Message, warm median)

MATCH (m:Message {id})-[:HAS_CREATOR]->(p:Person) RETURN p:

engine	full	fast	speedup	parity
pandas	40.6 ms	2.0 ms	20.3×	✅
polars	13.7 ms	3.44 ms	4.0×	✅
polars-gpu	24.6 ms	2.47 ms	10.0×	✅
cudf	27.8 ms	(falls through)	1.0×	✅

The fast path now engages on pandas/polars/polars-gpu; cuDF falls through (the
generic chain fast path already helps it).

Correctness

• Dispatch keys off the actual frame type (is_polars_df), not the requested
  engine — the WITH..MATCH reentry path can request engine=polars while handing
  a pandas-materialized intermediate graph, which would otherwise run polars ops
  on a pandas frame. (Caught by test_with_match_reentry_differential_fuzz.)
• Differential parity (fast vs full) byte-identical on polars, plus an
  independent oracle (hand-computed creators) and varlen-decline coverage.
• Zero non-seeded regressions: full compute/gfql + test_chain suite = 77
  failed / 5096 passed vs perf(gfql): seeded typed-hop fast path — seeded 1-hop RETURN in ≤1ms (#1755) #1759 baseline 77 / 5090 (identical failure set — all
  pre-existing local polars-drift/GPU-lib), +6 net new passing tests.
• mypy clean on both changed source files.

🤖 Generated with Claude Code

[lmeyerov]

Review responses:

• _seeded_typed_return_dst_polars direction is now the Direction Literal (was str), matching the pandas/cuDF helpers.
• CHANGELOG entry added.
• On location (lazy/engine/polars/ was floated): keeping it beside the other seeded helpers (moved to chain_fast_paths.py by the stacked refactor(gfql): extract seeded fast-path specializations to dedicated modules (#1755) #1762) rather than under gfql/lazy/engine/polars/ — it is a chain-level fast path dispatched from the cypher executor pre-pipeline, not part of the lazy row-pipeline engine; placing it in the lazy engine would create a back-edge from the engine into chain/AST internals and break the one-directional module-extraction convention (feat(gfql): add aggregate fast paths and final hardening #1731).