Implement CUDA QR fallback when cuQuantum is unavailable

[IvanaGyro]

Problem

NVIDIA supports cuTensorNet and cuStateVec on Linux only, so the Windows CUDA build in #1110 must use USE_CUQUANTUM=OFF. Most cuQuantum-gated paths already have another implementation: Gesvd_truncate and Rsvd fall back to the existing CUDA SVD path, and RegularNetwork falls back to Cytnx's contraction-order implementation. The clear remaining gap is linalg::Qr, which currently raises:

QR decomposition on GPU not implemented without cuQuantum support!

This makes a supported CUDA build lose a public operation solely because the optional cuQuantum backend is unavailable. The same gap is visible in the cuQuantum-disabled test configuration discussed in #1102.

NVIDIA support matrix: https://docs.nvidia.com/cuda/cuquantum/latest/getting-started/index.html#dependencies

Proposed scope

• implement a CUDA QR path that does not depend on cuQuantum, preferably using existing CUDA/cuSOLVER primitives;
• audit the other UNI_CUQUANTUM-gated public operations and document or fill any remaining functional gaps;
• preserve the cuQuantum implementation when USE_CUQUANTUM=ON;
• cover real and complex dtypes for both Tensor and applicable UniTensor entry points.

Acceptance criteria

• linalg::Qr works on CUDA tensors in a USE_CUDA=ON, USE_CUQUANTUM=OFF build;
• results satisfy reconstruction and orthogonality checks for float, double, complex float, and complex double;
• tests run in a cuQuantum-disabled CUDA configuration and no longer skip solely because cuQuantum is absent;
• behavior with cuQuantum enabled is unchanged.

Researched and authored by OpenAI Codex at @IvanaGyro's request while preparing #1110.