Add gated accelerator validation and benchmark matrix

[Tinnci]

Build a gated validation and benchmark strategy for accelerator support.

Scope:

• Add environment-gated tests for CUDA/ROCm/XPU/MPS/DirectML/ONNX providers where available.
• Record device info, provider info, sample rate, duration drift, peak/RMS, golden metrics, and runtime speed.
• Keep GitHub-hosted CI CPU-only unless a self-hosted runner is deliberately configured.
• Avoid turning throughput trend metrics into hard blockers until baselines are stable.

Acceptance criteria:

• Default test suite stays CPU/offline.
• Hardware tests are opt-in via environment variables.
• Benchmark output is machine-readable JSON.
• Docs explain how to run the matrix locally or on a self-hosted runner.