Give the two FFI example crates a runnable entry point

Beginne mit examples/tpch/run_tpch.py, den vorhandenen pytest-Dateien unter python/tests in beiden FFI crates und test.yml:136-150, um den Testbefehl zu bestätigen. Füge jedes Root run_demo.py, seinen subprocess-Treiber mit anchor assertions und den README Role-Header hinzu; fertig bedeutet, dass beide Skripte standalone ausgeführt werden und ihre Treibertests bestehen, wobei der zweite Decode-Abschnitt dem logical-codec sub-issue folgt.

Neither datafusion-ffi-example nor datafusion-ffi-query-planner-example has a script you can run. They are 51 and 47 pytest assertions respectively. Meanwhile user-guide/data-sources.md:219 calls the first one a "user example" and sends Python users there from the user guide, and io/table_provider.md:30 calls it "a complete example". We promise a runnable example and deliver a test suite.

Precedent: examples/tpch/ — the script is the artifact and a pytest keeps it honest. CI already runs pytest python/tests/_test*.py in each crate directory (test.yml:136-150), so a driver test needs zero workflow changes.

Layout per crate: run_demo.py at the tree root, parallel to the existing examples/distributed/run_tpch.py, and a driver at python/tests/_test_run_demo.py that runs it via subprocess.run([sys.executable, script]) and asserts on anchor lines. Subprocess rather than import_module or runpy, because it proves the thing being asked for: that it runs standalone with no pytest around it.

Each script should open with an import guard that turns the one real failure mode into an instruction rather than a traceback:

try:
    from datafusion_ffi_example import MyTableProvider
except ImportError:
    sys.exit("build the extension first:\n  uv run maturin develop\nSee README.md.")

• datafusion-ffi-example/run_demo.py — walk the conformance matrix in numbered sections: table provider, functions, catalog provider, config extension, codec round-trip, then the same bytes decoded a second time. That last section is only truthful after the logical-codec sub-issue lands, and it is the one place in the repository where the guide's central rule is demonstrated rather than asserted.
• datafusion-ffi-query-planner-example/run_demo.py — print plans, not rows. Planner nesting is understood by seeing the indented tree, and the existing tests assert on plan strings they never show. Cover the logical plan handed to the planner, the physical plan returned, the effect of SET ffi_query_planner.max_rows, and two planners nesting. Use .display_indent().
• Role: header on each tree's README, so the label survives the click from examples/README.md, which already distinguishes the trees but loses the distinction as soon as you follow the link. Three lines: what this tree is, where to go if you are learning the protocol, and how to run it.

Scope note: make them runnable, but do not promise datafusion-ffi-example reads as a tutorial. It is a coverage matrix — .ai/skills/check-upstream/SKILL.md:455 instructs adding every new FFI type to it — so it grows under protocol pressure rather than narrative need. The planner crate is single-topic and can carry the stronger claim.