feat(native): expose customer GPUOps through PyTorch
Add a shared pybind11 entry through customer torch_util, NodeState and OpKernel; preserve the dedicated HSTU native autograd interface. Include six existing HIP build compatibility fixes separately. Scope: 3 commits, 19 files, 1024 added lines. Validation: installed wheel on Torch 2.12 and HIP 6.3; 108 registered names passed supplemental forward checks, HSTU passed 6 output and gradient configurations, and RMSNorm JIT forward/backward passed with the existing external toolchain. Generic GPUOp entry is forward-only. Native static libraries were reused; full clean rebuild, Bazel targets and NVIDIA CUDA validation remain unverified. Target v1.2 remains unchanged until merge.