Add HIP/ROCm build paths and conditional GPU kernels for CUDA extensions

[Looong01]

Motivation

• Provide ROCm/HIP support across the repository so the CUDA extensions and kernels can be built for AMD/ROCm as well as NVIDIA/CUDA.
• Make backend selection explicit and controllable and ensure third-party bundled extensions (curope) follow the same backend logic.
• Avoid hard-coded CUDA-only APIs in kernel sources so code is portable between nvcc/cuda and hipcc/ROCm toolchains.

Description

• Added backend detection and selection to the top-level setup.py using SPLATT3R_GPU_BACKEND with auto|cuda|rocm|cpu options and emit backend-appropriate extra_compile_args and CUDAExtension configuration.
• Updated splatt3r_slam/backend/src/gn_kernels.cu and splatt3r_slam/backend/src/matching_kernels.cu to conditionally include hip/hip_runtime.h (and hip/hip_fp16.h where needed) under ROCm and to replace cuda::std usage with portable std::numeric_limits.
• Reworked both curope copies (splatt3r_core/src/.../curope and thirdparty/mast3r/.../curope) to add a CUROPE_GPU_BACKEND auto|cuda|rocm selector in their setup.py and to make kernels.cu use backend-agnostic aliases (gpuGetLastError, gpuGetErrorString, gpuError_t, gpuSuccess) guarded by USE_ROCM/__HIP_PLATFORM_AMD__.
• Documented ROCm install and build options in README.md and added examples for forcing builds (SPLATT3R_GPU_BACKEND and CUROPE_GPU_BACKEND).

Testing

• Ran Python syntax checks with python -m py_compile setup.py splatt3r_core/src/mast3r_src/dust3r/croco/models/curope/setup.py thirdparty/mast3r/dust3r/croco/models/curope/setup.py which completed successfully.
• Verified backend-selection validation by running CUROPE_GPU_BACKEND=invalid python .../curope/setup.py --name which produced the expected RuntimeError for an invalid backend.
• Confirmed source edits by searching for new ROCm symbols with rg (checks for hip_runtime, hip_fp16, CUROPE_GPU_BACKEND, PYTORCH_ROCM_ARCH, and numeric_limits<scalar_t>::lowest) which matched the updated files.
• Note: full compilation and runtime validation of the GPU extensions (CUDA or ROCm) could not be completed in this environment because the container lacks a full CUDA/ROCm toolchain and device drivers, so final build verification should be performed on a machine with the target GPU toolchain installed.

---

Codex Task