Fix the new interface for swap in cuda thrust operators.

[hzhangxyz]

Description

The new thrust interface requires std::swap or cuda::std::swap for thrust::sort, instead of the previous plain assignments, so we need to implement cuda::std::swap manually since cuda 12.9.

Closes: #36
See: NVIDIA/cccl#3333

Checklist:

• I have read the CONTRIBUTING.md.

---

给Reviewer看的简单介绍：

cuda 12.9 开始, thrust所在CCCL更新至3.0，有很多API break了，其中影响到我们的是，thrust的sort函数调用了swap函数，而std::array之前通过内部手动交换没有问题，但是他的swap并没有GPU实现，所以nvcc叫了。

解决方法：用cuda::std::array。

[hzhangxyz]

Wait, it seems this does not work.

[hzhangxyz]

I asked it there.

[hzhangxyz]

I tried another solution by replace std::array to plain array, this is ugly. Let's see if is there any other solution these days.

See branch: fix/cuda-12-9-error-about-thrust-sort-by-use-plain-array

Note: This branch does not work currently.

[hzhangxyz]

It seems we need to use cuda::std::array instead of std::array.

See: NVIDIA/cccl#5001 (comment)

[Copilot]

Pull Request Overview

This PR updates the array types in the CUDA Hamiltonian interface to use cuda::std::array instead of std::array so that the custom swap logic can be used with the new thrust requirements.

• Replaced array type references in helper structs and device functions.
• Updated reinterpret_cast calls in thrust sort functions to use cuda::std::array pointers.

Comments suppressed due to low confidence (2)

qmb/_hamiltonian_cuda.cu:245

• Verify that the layout of cuda::std::array is compatible with reinterpret_cast operations used in thrust sorting. If necessary, ensure a helper function or static_assert is in place to validate compatibility.

reinterpret_cast<cuda::std::array<std::uint8_t, n_qubytes>*>(sorted_result_configs.data_ptr())

qmb/_hamiltonian_cuda.cu:622

• Confirm that reinterpret_cast casts to cuda::std::array pointers remain safe and valid when used in thrust::sort calls, ensuring consistency with the new cuda interface.

reinterpret_cast<cuda::std::array<std::uint8_t, n_qubytes>*>(sorted_exclude_configs.data_ptr())

[hzhangxyz]

Implementing cuda::std::swap does not solve this issue, so close it.