Quarantine the greedy ForeignExecutionPlan claim so it cannot be copied by accident

Commencez par examples/datafusion-ffi-example/src/physical_extension_codec.rs:125-150 et les indications associées dans extension-guide/checklist.md. Suivez les statics et les helpers utilisés par MyPhysicalExtensionCodec, puis isolez le workaround derrière claims(), park() et take() dans src/foreign_plan_workaround.rs avec la documentation de module spécifiée. C’est terminé lorsque grep trouve exactement un “NOT A PATTERN” dans examples/ et que cargo clippy --no-deps --all-targets s’exécute sans erreur.

examples/datafusion-ffi-example/src/physical_extension_codec.rs:150 claims node.is::<DataSourceExec>() || node.is::<ForeignExecutionPlan>(). The second arm takes every other library's nodes, which extension-guide/checklist.md tells authors never to do.

This is not a bug to fix. It is G2 in #1719: the arm is load-bearing for upstream defect apache/datafusion#25152, and narrowing it to DataSourceExec alone makes 31 of the 51 tests in datafusion-ffi-query-planner-example fail. The object registry in the same file is downstream of it — once you claim a node you cannot introspect, there is nothing to write down about it, so parking it is all that is left.

The goal is to make it impossible to copy by accident while it stays. Proposal: move the statics and helpers into a new src/foreign_plan_workaround.rs behind three deliberately blunt functions — claims(), park(), take(). Nobody reads foreign_plan_workaround::park(node, buf) and thinks they are looking at serialization. The 25-line explanation currently buried inside a function body at lines 125-149 becomes the module //! doc, with three greppable fields:

//! # NOT A PATTERN
//! **Blocked on:** <https://github.com/apache/datafusion/issues/25152>
//! **Delete when:** `FFI_PlanProperties` carries `scheduling_type`, or
//!   `ForeignExecutionPlan` gains a reachable `try_to_proto`.
//! **Copying this will:** claim every other library's plan nodes, and produce
//!   payloads that decode only in the writing process, exactly once each.

One honesty note belongs in that doc: the DataSourceExec arm could be durable and is not, because the registry has to exist for the ForeignExecutionPlan arm regardless. Splitting the two arms across two wire formats costs real code and removes nothing. Saying so lets the guide claim the registry cannot be removed without also claiming every byte of it is forced.

Do not rename MyPhysicalExtensionCodec — 21 call sites plus two doc references, and the module doc already carries the label.

Done when: grep -rn "NOT A PATTERN" examples/ returns exactly one site, and cargo clippy --no-deps --all-targets is clean. No test changes expected; this is code motion plus documentation.