Add Python Torch conversions

[yuanknv]

Description

This pull request adds typed Python conversions between PyTorch tensors and tensor_msgs/ExperimentalTensor messages.

The new torch_conversions_py package exposes the torch_conversions Python module with the following APIs:

• allocate_tensor_msg(): Allocates CPU-backed or CUDA-backed message storage.
• from_output_tensor_msg(): Returns a writable Torch view over message-owned storage.
• from_input_tensor_msg(): Returns an independent tensor by default or a zero-copy read-only view with clone=False.
• to_tensor_msg(): Copies an existing Torch tensor into message-owned storage.
• set_stream(): Provides CUDA stream context management.

For CUDA-backed messages, a private pybind extension exposes the rosidl::Buffer allocation through DLPack. The DLPack context retains the buffer and its CUDA read or write handle until PyTorch releases the tensor, preserving asynchronous stream and event synchronization without copying the message storage.

This pull request is stacked on #6, which provides cuda_buffer_py.

Is this user-facing behavior change?

No. This pull request adds a new optional Python API without changing existing behavior.

This adds a new Python API for constructing ExperimentalTensor messages and accessing their CPU or CUDA storage as PyTorch tensors.

Did you use Generative AI?

Yes. OpenAI GPT-5.6 Sol via Cursor was used to assist with implementation and test.

Additional Information

This work extends the C++ Torch conversions introduced in #1 with an equivalent Python-facing API.

[nvcyc]

Do we need to add cuda_buffer and cuda_buffer_py as depndencies?

Suggested change

	<depend>rosidl_buffer</depend>
	<exec_depend>pytorch_vendor</exec_depend>
	<exec_depend>rosidl_buffer_py</exec_depend>
	<exec_depend>tensor_msgs</exec_depend>
	<depend>cuda_buffer</depend>
	<depend>rosidl_buffer</depend>
	<exec_depend>cuda_buffer_py</exec_depend>
	<exec_depend>pytorch_vendor</exec_depend>
	<exec_depend>rosidl_buffer_py</exec_depend>
	<exec_depend>tensor_msgs</exec_depend>