ONNX I/O module MVP: #230

• loader,
• decoders
• importer scaffold

Summary

Add a lightweight ONNX loader (using pbandk), float tensor decoding, and a proto-agnostic importer scaffold that maps ONNX graphs to ComputeGraph with initial support stubs for common ops.

• Why: Establishes the foundation for importing ONNX models into SKaiNET’s compute graph to enable wider model interoperability.

Acceptance Criteria

• Can load an ONNX model via OnnxLoader.fromModelSource and access both parsed proto and raw bytes.
• Float tensor initializers convert to internal tensors via TensorProto.toTensor().
• Importer constructs a ComputeGraph with:
  • Initializers surfaced as constant nodes.
  • Declared inputs/outputs wired into the graph.
• Unit tests validate loader and basic importer workflows.

Tasks

• Review/complete op mappings in OnnxImport.kt (ensure Conv, MatMul, Add, Mul, Relu, Sigmoid, Softmax, Reshape, Transpose, MaxPool, Upsample align with existing tensor ops and shape semantics).
• Expand dtype support (currently only FLOAT) and endianness handling; add tests for each dtype path.
• Add error reporting for missing/ambiguous inputs/attributes and unsupported ops (clear messages and remediation hints).
• Document how protobuf classes were generated and how to re-generate them via tools/.