Fix Tracr demo output_label: compiled labels are auto-numbered

[dewstend]

Description

The Tracr demo as shipped fails at its state-dict cell. Cell 11 passes output_label="reverse", but Tracr builds labels as f"{name}_{unique_id}", so the compiled expression's residual labels are reverse_1:* and make_tracr_transformer_bridge_state_dict raises:

ValueError: Could not find output basis label 'reverse:1' in model.residual_labels. Pass the final RASP expression label as output_label.

This is a follow-up to #1460 (the TransformerBridge port addressing #481): the conversion helpers there are correct, but the demo call site — and the unit-test fixture — assume the un-numbered label. The fixture in tests/unit/test_tracr_conversion.py hand-writes residual_labels containing "reverse:1" etc., a shape real Tracr never emits, which is why the tests pass while the demo doesn't run.

Three small changes, one concept (Tracr labels are auto-numbered; ask the expression for its label instead of hardcoding its name):

1. Demo cell 11: output_label="reverse" → output_label=reverse.label, plus a sentence in the preceding markdown cell explaining the auto-numbering (and that omitting output_label also works via infer_tracr_output_label).
2. Test fixture: "reverse:*" → "reverse_1:*" in _fake_tracr_model() and the matching output_label call sites, so the mock matches what the compiler actually produces. The ambiguity test still has two candidate prefixes (aggregate_1, reverse_1) and all positional assertions are unchanged.
3. Docstring: the infer_tracr_output_label example "reverse:3" → "reverse_1:3".

Fixes #481.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Validation

Run locally against this branch (Python 3.12, torch CPU, real tracr from google-deepmind/tracr — note the Tracr stack needs jax<0.11, since dm-haiku still references the removed jax.core.DropVar):

• pytest tests/unit/test_tracr_conversion.py — 5 passed.
• The demo's code path executed end-to-end as a script (compile the demo's RASP program → TransformerBridge.boot_native → load state dict → run): with output_label="reverse" it raises as above; with output_label=reverse.label and with output_label omitted, the Bridge decoding matches Tracr (['BOS', 3, 2, 1]) and every layer's attn/MLP hook_out matches Tracr's layer_outputs (np.isclose(...).all() per layer).

Prepared with AI assistance and human-reviewed; all validation runs above are real runs on the contributor's machine.