[RFC] Add support for megatron backend in Ascend NPU

[jiaqiw09]

This issue proposes adding optional Ascend NPU support to ROLL by integrating MindSpeed, which is essentially a toolkit to run megatron on ascend npu.

Motivation

On Ascend NPU, MindSpeed provides Megatron-compatible patches and runtime extensions.
ROLL currently does not expose a clean way to apply these patches or pass the required configuration, which makes Ascend execution difficult.

High-level approach

• Apply MindSpeed patches before Megatron modules are initialized
• Expose a small set of MindSpeed-required configuration fields at the config layer
• Allow MindSpeed-related arguments to pass through existing ROLL configs
• Keep all behavior opt-in and limited to Ascend NPU

Non-goals

• No behavior changes for CUDA / CPU
• No MindSpeed hard dependency
• No architectural changes to Megatron

Environment

• Target: Ascend NPU
• MindSpeed version: 0.14.0

[jiaqiw09]

@PanAndy I’m planning to open a WIP PR shortly. Feel free to comment here if you have feedback or preferences. I’ll coordinate with @UsernameFull as well.

If this direction looks reasonable, I can get a PR up pretty quickly.