[Feat]: add NPU fused operators (RMSNorm, RoPE, SwiGLU, SDPA)

[ys2025-AI]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Extends Twinkle's NPU support from a basic MoE GMM patch to a full fused-operator suite (RMSNorm, RoPE, SwiGLU, SDPA) for Ascend hardware.

Experiment results

Atlas 900 A2 (8× NPU) | Qwen3-30B-A3B-Instruct-2507 | LoRA r=8, batch=16, 188 steps | Dataset GSM8K_ZH

Metric	Baseline	This PR	Delta
Total	544 s	503 s	+7.5%
Training (step 10–180)	465 s	404 s	+13.1%
Loss / GradNorm	—	—	<< 0.01

[gemini-code-assist]

Code Review

This pull request introduces comprehensive NPU hardware acceleration support for Ascend devices by implementing fused operators (RMSNorm, RoPE, SwiGLU, and SDPA) and monkey-patching logic for specific model families like Qwen. It also refactors the NPU patching mechanism to be applied automatically when an NPU device is detected. Review feedback focuses on improving error handling by logging tracebacks for broad exception catches and restoring type hints and assertions that were removed during the refactoring of the MoE grouped matrix multiplication functions.

[ys2025-AI]

已改为在 kernel/init.py 中直接复用 Torch.is_npu_available()

[ys2025-AI]

已经通过pre-commit检查