[larissakl]

This PR prepares the ONNX module for a transformer state manager that keeps its KV cache states on the GPU. As long as all inputs and outputs are in host memory, Session::run uses the previous code path, so existing setups are not effected.

Onnx::Value can now hold tensors in CUDA device memory. New helpers check where a tensor is, copy it to the host or wrap existing device memory. Host accessors abort with a clear message when called on a device tensor.
Session::run gets an overload with a memory location per output. If any input or output is on the device, the run uses an Ort::IoBinding. By default, a state output stays on the device if its state input came from the device and everything else goes to the host. This way label scorers and LMs need no changes once a state manager provides device inputs.
All CUDA sessions of a process now share one CUDA stream, so other code can write device inputs on the same stream without extra synchronization. Note that each run synchronizes this stream, so sessions used from several threads wait for each other's GPU work. The new parameter cuda-use-tf32 allows disabling TF32. Requires onnxruntime >= 1.16.

Depends on #279.