fix(smollm3): resolve the NoPE schedule from the shared ModelConfig

[Moviw]

Background

Hi @ruiling-smartbear — huge thanks for the SmolLM3 family work; I'm not trying to duplicate it. While helping run the missing engine-build step on a real TensorRT box (RTX 3090, TensorRT 11.2.1.2), I hit a blocker in the builders and this is a small fix for it, offered against your feat/smollm3 branch so it can fold straight into NVIDIA#1158.

The engine builders call rope_layer_schedule() on the config object the production build pipeline hands them, which is the shared tensorrt_model_connect.config.ModelConfig (from engine_builder.py / build_cli.py), not the family-local copy where the method is defined. So both builders raise AttributeError: 'ModelConfig' object has no attribute 'rope_layer_schedule' before any engine is built.

What changed

• families/smollm3/config.py: moved the schedule resolution into a module-level rope_layer_schedule(config) helper that reads config.raw / config.num_hidden_layers and therefore works on the shared ModelConfig too; the ModelConfig method now delegates to it.
• standard_decoder_builder.py and dual_profile_decoder_builder.py: call the module-level helper instead of the family-local method.
• build_routing.py: call the helper directly instead of getattr(config, "rope_layer_schedule", None), so malformed-schedule validation actually runs instead of silently no-oping.

No shared infrastructure is touched; all changes stay under families/smollm3/.

Validation

python3 tools/model_ci.py validate                                              # passes: 93 families, smollm3 present
pytest tests/e2e/models/smollm3/test_smollm3_standard_decoder.py -q             # 12 passed (was 11 failed / 1 passed)
pytest tests/e2e/models/smollm3/test_smollm3_builder_engine.py -q               # 16 passed (was 6 failed / 10 passed)
pytest tests/e2e/models/llama/test_llama_standard_decoder.py -q                 # 12 passed (control, same box)

The engine-building failures above are the same six the PR describes as "need a real TensorRT module" — on a TRT box they were not skipped, they were failing with the AttributeError above, and now pass.

Notes

• Base: ruiling-smartbear:feat/smollm3 (this PR targets your branch, not NVIDIA:main).
• The full test_model_e2e[smollm3-3b] parity run still needs the 3B checkpoint + engine build; I did not run it (needs /mnt/storage + model download), which is unchanged from your remaining gap.

Refs: NVIDIA#1146, NVIDIA#1158