Support SmolLM shape tensors in CoreML

[matthargett]

Summary

• lower internal shape nodes to MIL shape
• infer and preserve shape-tensor rank and type through imported unsqueeze nodes
• cast integer where conditions to uniquely named MIL bool values
• verify exact CoreML predictions for shape -> unsqueeze (int64 and uint32 readback) and zero/nonzero integer masks

This completes the first three items in #222. The SmolLM compiler pass now reaches the dynamic attention_mask input-signature mismatch tracked there.

Validation

• M4, Rust 1.97.0: make fmt-check, make lint, and strict Clippy with coreml-runtime,dynamic-inputs pass
• Rust library tests: 357 default; 364 with coreml-runtime; 371 with coreml-runtime,dynamic-inputs; all pass
• TensorRT option-operand consumer regression test passes
• full CoreML WPT at 04cd4a830e26add75d385f28d95a080600b5d70f: 2,460 passed, 41 expected skips, 20 existing negative-index scatter failures. Results match a clean main control case-for-case
• make docs-backend-ops-check passes
• earlier M4 SmolLM-135M validation: compilation advances from unsupported shape and duplicate MIL values to Model input 'attention_mask' has a different shape than its corresponding parameter to main.

The public builder/interchange shape dtype remains int64; the uint32 test uses an explicit cast. This does not implement #225's proposed dynamic-shape API. A reduced ranged shape -> unsqueeze probe still reproduces the input-signature mismatch tracked in #222; the passing numerical tests use fixed dimensions and repeated CPU-only predictions.