Replace the logical codec's object registry with durable metadata

Commencez dans examples/datafusion-ffi-example/src/logical_extension_codec.rs, au niveau de try_encode_table_provider, et comparez l’approche IPC dans examples/distributed/storage-library/src/codec.rs. Mettez à jour python/tests/_test_logical_extension_codec.py, notamment l’assertion concernant l’ancien codec et un test de décodage deux fois. C’est terminé lorsque le registre et le code de token ont disparu, que provider_prefix est utilisé, que les tests spécifiés passent et que grep -in token ne renvoie rien.

examples/datafusion-ffi-example/src/logical_extension_codec.rs parks live table providers in a process-global HashMap and encodes an integer token into it. Encoding inserts, decoding removes, so the same bytes cannot be decoded twice, one plan cannot fan out to several readers, and a plan that never reaches a decoder keeps its provider alive for the life of the process. extension-guide/codecs.md tells authors not to do this.

Unlike the physical codec in the same crate (see the quarantine sub-issue), this one is fixable: try_encode_table_provider at line 148 claims node.downcast_ref::<MemTable>(), which is narrow, and a MemTable is fully describable by its schema and batches.

Pattern to copy: examples/distributed/storage-library/src/codec.rs — same Arrow IPC technique, same error convention (internal_datafusion_err! on encode, since this process holds the object; exec_datafusion_err! on decode, since those are foreign bytes).

Verified prerequisites: MemTable.batches is pub (datafusion-catalog/src/memory/table.rs:69), typed Vec<PartitionData> where PartitionData = Arc<tokio::sync::RwLock<Vec<RecordBatch>>>. MemTable::try_new rejects zero partitions (table.rs:84). arrow is already a dependency with IPC available, so no Cargo.toml change.

Proposed wire format, keeping the per-instance prefix the dispatch tests rely on:

<provider_prefix> | b"MEMTBL1" | u32 LE n_partitions | { u32 LE ipc_len | ipc stream }*

One stream per partition, because MemTable partition boundaries become output partitions. A stream carries its schema even when empty, so an empty partition round-trips.

Two traps worth writing down before someone hits them:

• Use try_read() on each partition lock, not blocking_read(). The FFI codec runs with a tokio runtime handle installed, and blocking_read panics in that context.
• On decode, build MemTable::try_new from the IPC schema, not the schema: SchemaRef argument. try_new validates schema.contains(&batch.schema()), so metadata drift would surface as a spurious mismatch. This is the opposite choice from storage-library/src/codec.rs:376-380, which must honour the plan's schema because it re-reads files from disk; here the batches are the payload. Worth a comment noting the contrast, since the two codecs otherwise look alike.

Done when: the registry, token_id(), and the HashMap/Mutex/OnceLock/AtomicU64 imports are gone; the struct field token is renamed provider_prefix to match the Python kwarg that already uses that name; and grep -in token over the file returns nothing.

Tests: of 19 tests in python/tests/_test_logical_extension_codec.py, one changes. test_installing_a_codec_cannot_hijack_an_earlier_codecs_objects asserts len(before) == len(after) with a comment about tokens being minted per encode; that comment becomes false and the assertion becomes weaker than reality, so it should become assert before == after. Add one test for the property the guide claims and nothing currently covers: encode once, decode twice on one session, assert both produce the same rows. All 47 tests in the planner crate should be unaffected — every assertion there is on call counters, never on payload shape.