Raw FFI + dynamic loader for NVIDIA TensorRT (C API surface).
TensorRT's native public API is C++; NVIDIA ships a partial C-ABI surface
suitable for language bindings in NvInferRuntimeCAPI.h (TRT 10+). This
crate wraps that surface for runtime deserialization and inference. The
builder side of TensorRT remains C++-only; use the TRT trtexec tool or
the Python bindings to produce serialized engines, then load them here.