[Feature][QDP] Add AMD GPU support via Triton backend

[400Ping]

Related Issues

Closes #1155

Changes

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other

Why

QDP currently assumes the CUDA/native _qdp path, which makes the Python API effectively NVIDIA-only.
This change adds an AMD/ROCm execution path so QDP can run on AMD GPUs while keeping the existing CUDA route intact.

The goal of this PR is to:

• reduce NVIDIA vendor lock-in for QDP users
• keep a single QdpEngine entry point for backend selection
• document and test the AMD/ROCm route without regressing the existing CUDA/native path

User-Facing Behavior

CUDA:

from qumat.qdp import QdpEngine

engine = QdpEngine(device_id=0, backend="cuda")
qtensor = engine.encode(data, num_qubits=2, encoding_method="amplitude")

AMD/ROCm:

from qumat.qdp import QdpEngine

engine = QdpEngine(device_id=0, backend="amd", precision="float32")
tensor = engine.encode(data, num_qubits=2, encoding_method="amplitude")

Testing

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes

[400Ping]

Just noticed that colab doesn't have AMD GPU options, so we may need to ask AMD for collaboration. Or I will looked into modal if I can rent one.

[400Ping]

On second thought, let me see if I can rent one first.

[guan404ming]

Let's turn this one to draft to prevent early review.

[400Ping]

# python qdp/qdp-python/benchmark/benchmark_triton_amd.py --qubits 12 --batch-size 64 --batches 200 --encoding-method amplitude
TRITON AMD BASELINE
- Encoding: amplitude
- Qubits: 12
- Batch size: 64
- Batches: 200
- Duration: 0.0153 s
- Throughput: 835843.1 vectors/sec
- Latency: 0.001196 ms/vector

Tested on AMD Instinct MI300X

[400Ping]

cc @viiccwen to take a look

[guan404ming]

Thanks for the patch, just left some nit comments.

[guan404ming]

Great, I think we could work on this and refine this. Thanks!

[ryankert01]

This PR introduce two ways for amd GPUs to run qdp, could be refine later.

Could be a followup.

[rich7420]

thanks for the patch! left some comments

[400Ping]

I am fixing some of the implementations, putting this to draft.