[Feature] Improve usability on Ascend NPUs

[HwVanICI]

Checklist

• This feature will maintain backward compatibility with the current APIs in
  areal/api/. If not, please raise a refactor issue first.

Background

Currently, AReaL supports RL training on Ascend NPUs, but it's not very user-friendly out of the box (see #505). We plan to address this with a series of steps:

1. Keep an ascend branch inside the official AReaL repo. It will have minimal adaptation code that only affect NPU training and should not be merged to the main branch for GPU usage. Before step 3 is done, this branch will be maintained to minimize its gap with the main branch through frequent rebasing to the main branch.
2. Provide documentation for how to install AReaL on NPU platforms, where we will also update the NPU docker image accordingly. Other NPU documentation will be rolled out in the future.
3. Refactor the NPU-specific code of step 1 gradually to make them compatible with GPU code, and send PR for merging into main branch (including fixing [Feature] Add an npu section in pyproject.toml #508).