[Feature Request] Add Ascend NPU Native Support for TorchRL

[lowdy1]

Motivation

Currently, TorchRL primarily supports CPU and CUDA backends. We would love to see native support for Ascend NPUs, so that researchers and developers working in NPU-based ecosystems can run reinforcement learning workloads smoothly and without extra workarounds.

Solution

We plan to validate this step by step through example experiments, starting with single-agent setups, then extending to multi-agent and distributed algorithms. We will also write tests and update documentation along the way. During this process, we will also take the opportunity to fix issues we encounter in the repository and contribute improvements back. Along the way, we are happy to open PRs that bring native NPU support into TorchRL.

Checklist

• I have checked that there is no similar issue in the repo (required)