📝 [NPU] Use vendor-neutral wording in quantization comments

[TallMessiWu]

Motivation

Follow-up to #34829. That PR cleaned the MXFP8/MXFP4 linear methods, but the
NPU quantization comments still carried two kinds of vendor wording, all landed
by my earlier NPU quantization PRs:

1. Comments naming external projects, left in five files 📝 [NPU] Clean up quantization comments #34829 did not reach.
2. Ascend used as prose branding in docstrings for modules that are otherwise
   named and namespaced npu.

Modifications

Comment-only cleanup, no code touched.

Third-party references removed:

• srt/hardware_backend/npu/moe/init_routing.py, .../quantization/moe_methods.py,
  srt/models/qwen3_moe.py — the MXFP8 MoE scale-normalisation docstring, the
  fused-MoE weight-layout note and the Qwen3 MoE router-gate note (from ✨ [llm][npu][quant] Add W8A8 MXFP8 quantization for Qwen3 MoE on Ascend NPU #30768).
• srt/hardware_backend/npu/quantization/linear_method_npu.py — the
  NPUDualLevelMXFP4LinearMethod docstring (from ✨ [llm][npu][quant] Add W4A4 MXFP4 quantization support for Qwen3 Dense on Ascend NPU #23795).
• multimodal_gen/runtime/layers/quantization/modelslim_mxfp4_scheme.py and
  mxfp4_npu.py — the offline ModelSlim MXFP4 scheme and the online diffusion
  MXFP4 method (from ✨ [diffusion][npu][quant] Add MXFP4 quantization support for Wan2.2 Diffusion on Ascend NPU #22338).

Every removed line was pure attribution: the surrounding comments already state
the layout, format and calibration requirements those references pointed at
(FRACTAL_NZ format 29, the [in/level0_block_size, out] L0 scale transpose, the
smooth-quant mul_scale requirement, the strided transpose views). Where a
reference carried real information — the online MXFP4 diffusion weight path being
experimental because the established path is offline-only — the fact is kept and
only the project name dropped.

Ascend prose reworded to NPU:

• linear_method_npu.py — six class docstring headers, plus the three hardware
  requirement lines (Ascend 950 (A5) → A5 NPU; the file already uses the bare
  A5 / A2/A3 shorthand elsewhere).
• moe_methods.py — the two dispatcher dtype notes and the NPUMXFP8MoEMethod
  header.
• online_moe_methods.py, modelslim_mxfp4_scheme.py, modelslim_mxfp8_scheme.py,
  mxfp4_npu.py — module and class docstring headers.

Identifiers are deliberately untouched, since they are real API names and CLI
values: AscendQuantInfo, AscendRunnerCore, MoeRunnerBackend.ASCEND,
is_ascend_fuseep(), ascend_tp / ascend_fuseep, and the ascend attention
backend. The two runtime error strings that name the hardware are also left as
they are, so user-facing messages do not change.

Accuracy Tests

Not applicable — comments only, no behaviour change. python -m black --check
and an AST parse pass on all eight files.

Speed Tests and Profiling

Not applicable — comments only.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #33155197134
Latest PR Test (Extra): ❌ Run #33155196857
Latest PR Test (AMD ROCm 7.2): ❌ Run #33155196931

[ping1jing2]

i merged it without running CI as there is no code change in this PR