Define accelerator capability model and fallback policy

[Tinnci]

Define the cross-backend accelerator model before adding provider-specific code.

Scope:

• Separate logical device selection (cpu, cuda, rocm, xpu, mps, directml, auto) from backend implementation (self_torch, self_onnx, external_package, external_sdk).
• Extend capability metadata without making every backend import heavy runtimes at listing time.
• Define fallback behavior when a requested accelerator is unavailable.
• Keep default CPU/offline behavior deterministic.

Acceptance criteria:

• --list-models / --list-backends can report accelerator support clearly.
• device=auto behavior is documented and testable.
• Unsupported requested accelerators fail with actionable errors.