feat: support ascend npu sft

[donpromax]

Summary

This PR adds Ascend/NPU support for SFT workloads in StepTronOSS.

It introduces a NPU runtime patch entrypoint, NPU-specific optimization backends for MoE dispatch and grouped GEMM, Ascend SFT experiment entrypoints,
benchmark scripts, tests, and user-facing documentation.

Main Changes

• add manual NPU runtime patching through steptronoss.utils.npu_patch.apply_npu_patch()
• add grouped_gemm="npu_gmm" backed by MindSpeed npu_gmm_v2
• add TokenDispatcher="npu_alltoall" for EP token routing on NPU
• patch several CUDA-specific helper paths to work on NPU runtime
• register NPU backends when apply_npu_patch() is called
• add Ascend SFT experiment entrypoints for:
  • Qwen3-1.7B
  • Qwen3-30A3B
  • Step3 toy SFT
  • Step3.5-Flash Midtrain SFT with Muon
• add a real-data preparation script for Qwen3 SFT smoke and validation runs
• add NPU benchmarks and test cases for dispatcher and grouped GEMM
• add user documentation:
  • docs/ASCEND.md
  • docs/ASCEND_ZH.md

Implementation Notes

• Ascend patching is now explicitly enabled by calling apply_npu_patch() before importing modules that depend on NPU runtime behavior or before selecting NPU backends
• npu_alltoall builds a unique token-rank routing layout before communication to avoid duplicated routing to the same remote rank
• the fast path uses MindSpeed npu_moe_token_permute / npu_moe_token_unpermute on the NPU bf16 path
• the dispatcher keeps an index_select / index_add fallback path for compatibility
• npu_gmm keeps the same semantic interface as the existing grouped_gemm path and normalizes batch_sizes internally

Tested Environment

Validated with:

• CANN 8.3.RC2
• torch_npu 2.8
• MindSpeed v2.2.0_core_r0.12.1

All other dependencies are kept aligned with the project's uv environment.

Validation

• added dispatcher unit tests in tests/test_npu_alltoall_dispatcher.py
• added grouped GEMM correctness test in tests/test_grouped_gemm_npu.py
• added microbenchmarks:
  • benchmarks/benchmark_dispatcher_npu.py
  • benchmarks/benchmark_grouped_gemm_npu.py
• ran end-to-end SFT validation on representative Qwen3 and Step3.5 configs

Training Results

Qwen3-1.7B sft smoke test (compared to l40s):

Step-3.5-Flash-Base-Midtrain sft: tp8pp8vpp3ep8 seq_len=8192 with recompute, sequence_parallel and offload_optimizer_state

Kernel Optimization

NPU grouped GEMM speedup vs baseline:

[npu_grouped_gemm] name=moe_like_large group=36 batch=3256 k=4096 n=2560 dtype=bf16 trans_b=True
backend, fw_ms, bw_ms, total_ms
baseline, 11.976, 235.589, 247.565
npu_gmm_v2, 8.257, 19.858, 28.115
speedup_vs_baseline, fw=1.45x, bw=11.86x, total=8.81x
metric, close, max_abs_diff
forward, True, 0.000000
forward_bw_run, True, 0.000000
grad_a, True, 0.000000
grad_b, True, 0.000000

NPU MoE AllToAll Dispatcher speedup vs baseline:

[Randomizez]

感觉有很多可以简化的地方，请看一下。另外alternatives尽量直接注册，在patch内注册有点太过隐式了。

[Randomizez]

LGTM