feat: add npu dockerfile and useage

[UsernameFull]

PR: feat: add NPU Dockerfile and Ascend documentation

Summary

Add Docker build files and documentation for running ROLL on Huawei Ascend NPU (Atlas 900 A2/A3 PODc), enabling users to deploy RLVR and other pipelines on Ascend hardware.

Changes

Docker

• Add Dockerfile.A2 for Atlas 900 A2 PODc (Ascend 910B1, CANN 8.5.1)
• Add Dockerfile.A3 for Atlas 900 A3 PODc (Ascend 910_9391, CANN 8.5.1)
• Both images include: PyTorch 2.8.0+cpu, vLLM 0.13.0, vLLM-Ascend 0.13.0, DeepSpeed 0.16.4, Transformers 4.57.6, triton-ascend 3.2.0

Documentation (English + Chinese)

• Add ascend_docker_usage.md — Docker image build/pull, container startup, environment verification, and RLVR pipeline example
• Update ascend_usage.md — Revise Ascend support status and installation instructions

Example Config

• Add examples/ascend_examples/qwen3_8b_rlvr_deepspeed.yaml — Qwen3-8B RLVR config using DeepSpeed ZeRO-3 + CPU offloading on NPU

Files Changed (7 files, +808 / -68)

File	Change
docker/Dockerfile.A2	New
docker/Dockerfile.A3	New
docs_roll/docs/User Guides/Hardware Support/ascend_docker_usage.md	New
docs_roll/docs/User Guides/Hardware Support/ascend_usage.md	Updated
docs_roll/i18n/.../ascend_docker_usage.md	New (Chinese)
docs_roll/i18n/.../ascend_usage.md	Updated (Chinese)
examples/ascend_examples/qwen3_8b_rlvr_deepspeed.yaml	New

Key Notes

• NPU does not support Megatron-LM; DeepSpeed is the only supported training backend
• NPU does not support colocated mode; training and inference must use separate NPU devices
• flash_attn package is not supported; use fa2 via transformers instead