fix(bridge): restore native stop_at_layer and input_to_embed

[emerardd]

Description

Fixes #1632.

Native TransformerBridge blocks are named layers.N, but the block layer-index parser only recognized layers.N when it had a preceding dot. As a result, direct forward(..., stop_at_layer=...) calls never stopped and input_to_embed() returned final vocabulary logits instead of the residual stream entering block 0.

This change:

• makes the layer-name parser boundary-aware for blocks.N, h.N, and layers.N, including top-level native names;
• normalizes negative stop_at_layer indices before passing the stop point to blocks;
• lets the native model accept inputs_embeds, which is required to resume a residual with start_at_layer;
• adds direct stopping and input_to_embed() round-trip regression coverage for an offline boot_native() model.

No new dependencies are required.

Type of change

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

Screenshots

Not applicable.

Verification

• Before the fix, the four new regression cases failed: direct stops returned logits with the d_vocab dimension, and residual re-entry raised because the native model required input_ids.
• uv run pytest tests/unit/model_bridge/test_boot_native.py — 35 passed
• Focused direct-stop and round-trip regression selection — 4 passed, 31 deselected
• uv run pytest tests/integration/model_bridge/test_bridge_input_to_embed.py tests/integration/model_bridge/test_bridge_stop_at_layer.py::test_manual_hooks_stop_at_layer_compat_with_processing — 5 passed
• uv run mypy . — success, 424 source files
• pycln, isort, and black checks on the four changed files — passed
• git diff --check — passed

The complete unit-test suite was not rerun; only the affected unit and integration surface was run locally.

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation (not applicable; behavior now matches the documented API contract)
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes (only the affected test surface was run)
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

Thanks for filing and resolving this issue @emerardd, looks great, merging as is.