Add a reference ONNX post-processing pipeline for both Vision v8 branches

[Complexity-ML]

Context

Vision v8 ONNX exports expose raw tensors shaped [batch, 34000, 148]: 68 LTRB/DFL regression logits followed by 80 quality-class logits.

Consumers currently need to implement the deployment contract themselves:

• O2M: decode plus NMS.
• NMS-free: decode plus confidence filtering.

A shared reference implementation would reduce integration errors and make end-to-end benchmarking meaningful.

Scope

• Add a small framework module and CLI example that performs preprocessing, ONNX Runtime inference, decode, and post-processing.
• Support both exported branches from their metadata sidecars.
• Implement DFL box decoding once and share it between branches.
• Apply configurable confidence filtering to both branches.
• Apply class-aware NMS only for O2M.
• Return a stable detection schema with normalized and pixel-space boxes, class IDs, and scores.
• Measure model latency and post-processing latency separately.

Acceptance criteria

• Synthetic unit tests cover grid mapping, DFL decode, confidence filtering, and O2M NMS.
• End-to-end outputs are compared against the PyTorch detector post-processing on fixed inputs.
• NMS-free execution never invokes NMS.
• The CLI accepts provider selection and reports the provider actually used.
• Documentation includes runnable CPU and CUDA examples.

Related: #10