Add runtime provider abstraction for optimized execution

[Tinnci]

Add a runtime-provider layer that sits below model backends and above device detection.

Candidate providers:

• PyTorch eager / torch.compile where useful.
• ONNX Runtime CPU, CUDA, ROCm, DirectML, OpenVINO providers.
• TensorRT or OpenVINO SDK adapters where model export and licensing allow it.
• Intel XPU / IPEX route if it can be tested without contaminating default installs.

Rules:

• Backends request a runtime capability; they do not own provider installation or global detection.
• Provider code must be optional and import-light until selected.
• Provider selection must be explicit or driven by documented auto policy.

Acceptance criteria:

• Proposed API for runtime provider selection is documented.
• At least one provider adapter has tests with mocked availability.
• Backend code remains provider-agnostic unless a model genuinely needs a provider-specific graph.