docs(gfql): make every quick-reference example run on pandas and Polars (#2169) - #2170

Closes #2169. Keeps #967 (restoring local callables) open.

Problem

The GFQL cheatsheet recommended n({"age": lambda x: x > 30}), which validation rejects on pandas and Polars with GFQLTypeError E201. The docs example audit missed it for two reasons:

• its RST regex only matched .. code-block:: at column zero, so blocks nested under bullets were never extracted: 50 of 56 in quick.rst, 58 of 407 across GFQL docs;
• bare matchers like n({...}) only construct an AST and never execute.

Changes

• quick.rst: declarative predicates (gt, lt, between) replace callables; let() examples select bindings with output= (results never had per-binding columns); remote() shown as a let() binding, since a plain-chain remote step raises NotImplementedError; pandas-only .empty replaced with len(...); the engine section states that engine='auto' runs Polars-unsupported features (query= strings, same-path where=) on pandas and returns pandas frames, while engine='polars' raises; stale xfail removed; placeholder and server-dependent blocks marked doc-test: skip.
• docs/test_doc_examples.py: indent-aware RST extractor. On every other GFQL doc it yields a strict superset of the old blocks with identical markers. New test executes every runnable quick.rst example on a pandas-bound and a Polars-bound graph and requires identical rows, with explicit expected rows for the predicate examples. A pandas result under auto is accepted only when explicit engine='polars' refuses the same matcher.
• graphistry/tests/compute/test_callable_filter_rejection.py: pins structured E201 rejection of callables in node filters, edge matches and destination matches, on both engines.

No runtime or serialization change.

Validation (dgx-spark, image test-rapids-official:26.02-gfql-polars, polars 1.35.2)

• pytest docs/test_doc_examples.py: 24 passed, 13 skipped. 31 quick.rst examples execute on both engines with matching rows.
• Negative control: restoring the lambda example fails the new test on both engines with E201.
• test_callable_filter_rejection.py: 6 passed.
• Remote let() examples reach Must call login() first, not NotImplementedError.
• ruff clean; comment-density and type-hygiene guards OK.

Limits: the first Cypher example (:Person/:FOLLOWS) runs with row parity but returns no rows on the shared fixture. GPU, remote and compute_cugraph examples stay auto-skipped. The same wrong let() access pattern also appears in skip-marked blocks in overview.rst, about.rst and the spec docs; that is tracked separately.

🤖 Generated with Claude Code