perf(gfql): native polars translation for simple connected-join residuals (#1729/#1755)

[lmeyerov]

What

The connected-join grouped-count fast path applies scalar residuals — the shapes
#1729's lowering cannot push into filter_dict (toLower(a.col) = toLower('lit'),
a.col <op> literal) — by spinning a where_rows chain per alias (~1.7 ms each).
On graph-bench q5/q6/q7 that residual-apply was ~5 ms of a ~10 ms query.

On polars, translate exactly those simple shapes directly to native polars filter
expressions (_residual_polars_expr). Any expression outside the recognized
shapes falls back to the existing where_rows chain evaluator, so semantics can
never diverge.

Measured (dgx, 20k graph-bench, polars, warm median)

query	before	after
q5	10.2 ms	6.6 ms
q6	10.7 ms	8.9 ms
q7	13.5 ms	10.5 ms

Values unchanged (q5=1111 etc.).

Byte-parity verification

• Differential fast-vs-fallback identical on 8 adversarial cases: nulls, case
  variants, unicode casefold (Straße), numeric ranges, multi-expr conjunction.
• Full cypher lowering + row-pipeline suites: 1645 pass — including the
  residual-specific connected-join tests that caught an earlier (rejected)
  implementation approach.

🤖 Generated with Claude Code

[lmeyerov]

Review responses:

• _residual_polars_expr return type is now Optional['pl.Expr'] (TYPE_CHECKING import), was Optional[Any].
• expr: str is the honest input type — the feat(gfql): plan connected join predicate pushdown #1729 lowering serializes residuals as canonical predicate strings in ASTCall params (a typed term needs a lowering-level refactor, out of scope); rationale now in the docstring.
• CHANGELOG entry added.
• New committed regression suite test_residual_polars_native.py (20 tests): positive shapes incl. nulls/casefold/negative literals, negative declines, the all-or-nothing group-fallback gate, pandas-frames-never-fast-lane.