Document accelerator install strategy and optional extras

[Tinnci]

Document accelerator installation paths without making heavyweight SDKs baseline dependencies.

Scope:

• Decide which accelerators can be Python extras and which must be documented install commands due to vendor wheels/indexes.
• Cover PyTorch CUDA/ROCm, Intel XPU/IPEX, ONNX Runtime providers, DirectML, OpenVINO, and TensorRT at a planning level.
• Keep audio-super-resolution baseline install CPU-light.
• Add clear compatibility notes for Python versions and Windows/Linux differences.

Acceptance criteria:

• README/docs explain supported and experimental accelerator install paths.
• pyproject.toml extras are added only where packaging is stable and cross-platform enough.
• Missing SDK errors point users to the right install section.