Summary

Add an AMD GPU backend using Triton on ROCm so QDP can run efficiently on AMD hardware, not only NVIDIA CUDA GPUs.

Primary Motivation

We want to avoid NVIDIA vendor lock-in.
Supporting AMD GPUs gives us a second production GPU path and improves portability, procurement flexibility, and long-term maintainability.

Scope

• Implement a triton_amd backend for:
  • amplitude encoding
  • angle encoding
  • basis encoding
• Integrate backend routing and runtime capability checks.
• Add correctness parity tests against CUDA/JAX reference outputs.
• Add baseline benchmarks for AMD backend performance.
• Add docs for ROCm/Triton setup and usage.