This commit replaces InfiniTensor's per-device kernel implementations with InfiniOps, a unified operator library shared across the InfiniTensor ecosystem. This is the first step in a multi-part migration.

• Add InfiniOps as a git submodule at 3rd-party/infiniops, version-pinned via .gitmodules. The submodule provides kernel implementations for CPU, CUDA, and other accelerators through a single libinfiniops.so.

• Introduce the InfiniOps bridge layer (include/core/infiniops_bridge/):

adapter_kernel.h — base class for adapter kernels that delegate to InfiniOps C++ APIs.

tensor_convert.h — converts InfiniTensor TensorObj (shape, dtype, strides, device) to InfiniOps Tensor as a non-owning view.

infiniops_runtime.h / infiniops_runtime.cc — unified InfiniOpsRuntimeObj that replaces all per-device runtime classes (CudaRuntimeObj, BangRuntimeObj, KunlunRuntimeObj, etc.) with device-specific alloc/dealloc dispatching and workspace management.

cpu_fallback.h — placeholder for future GPU→CPU fallback path for operators not yet supported by InfiniOps on GPU.

• Implement 6 InfiniOps adapter kernels (src/kernels/infiniops/): Add, Mul, MatMul (with bias fusion via Gemm beta=1), Cast, Concat, and RMSNorm. Each adapter converts InfiniTensor tensors to InfiniOps tensors and calls the corresponding infini::ops::*::Call().

• Remove legacy per-device kernel implementations for CUDA, BANG, Kunlun, Ascend, and IntelCPU backends (~24,700 lines deleted). This includes all kernel source files, runtime classes, operator timers, and device-specific headers. The remaining CPU kernels (src/kernels/cpu/) serve as fallback for operators not yet covered by InfiniOps adapters.

• Simplify the build system: remove USE_CUDA, USE_BANG, USE_KUNLUN, USE_ASCEND, USE_INTELCPU CMake options. InfiniOps device backends are now controlled through InfiniOps's own WITH_CPU/WITH_NVIDIA etc. flags. GPU support is re-enabled via -DWITH_NVIDIA=ON passed through to InfiniOps's add_subdirectory.

• Update the Python FFI (src/ffi/ffi_infinitensor.cc): replace all device-specific runtime factory functions (cuda_runtime(), bang_runtime(), etc.) with a unified cpu_runtime() / cuda_runtime() that returns InfiniOpsRuntimeObj. Fix copyout_numpy to use numpy.empty() instead of py::array(dtype, shape, nullptr) to avoid NumPy 2.x stride issues.