Emit initializers as static const arrays and add golden/feature tests

[strimo378]

Motivation

• Treat ONNX initializers as true constants so model parameters can be embedded in generated C as read-only arrays.
• Produce deterministic, minimal generated C by emitting weights as static const arrays rather than runtime inputs.
• Allow the Python runtime to resolve initializer-backed tensors transparently when running compiled models.
• Add golden and end-to-end feature tests to cover initializer-backed weight emission and behavior.

Description

• Add Initializer to the IR and parse ONNX initializers in import_onnx, excluding them from the runtime inputs and storing them on the Graph as initializers.
• Introduce ConstTensor and add a constants field to LoweredModel, with _lowered_constants in compiler to lower Initializers into ConstTensors.
• Implement CEmitter._emit_constants and CEmitter._format_float to render initializer data as static const float <name>[...] = { ... }; and include the constants section in emitted C before operator functions.
• Resolve constants at runtime by adding _ResolvedFeeds.fetch and switching Compiler.run/internal run paths to use resolved feeds so initializer arrays are returned automatically.
• Rework lowered model input collection to use graph inputs, and propagate constants into matmul/binary/unary lowering paths.
• Add tests and fixtures: new tests/test_endtoend_features.py, added a golden file tests/golden/add_initializer_model.c, updated tests/test_golden.py, and adjusted tests/test_endtoend_ops.py to remove the duplicated initializer end-to-end test.

Testing

• Ran the full test suite with pytest -n auto -q and all tests passed (20 passed).
• Compiled and wrote the new golden C fixture via the in-repo Compiler to tests/golden/add_initializer_model.c during the rollout.
• Executed the new end-to-end feature test which compiles the generated C to an executable and compares outputs against ONNX Runtime.
• No automated tests failed during validation.

---

Codex Task