feat: support Qwen3.5 eager-mode on NPU with Python executor.

[Wang-1F]

Summary

Support Qwen3.5 on NPU via the Python executor path by replacing all PyTorch small-op kernel stubs with high-performance NPU operator calls.

What this PR does

• Registers 8 new torch.ops.xllm_ops operators for the Python executor's GDN (Gated Delta Net) layer on NPU:

  • rms_norm_gated — fused RMSNorm + SiLU gate (CANN aclnnLayerNormFwd)
  • l2_norm — L2 normalization (Triton-on-Ascend)
  • causal_conv1d_update — decode-path causal convolution (TileLang)
  • causal_conv1d_prefill — prefill-path causal convolution (CANN aclnnCausalConv1d)
  • causal_conv1d_qkv_prefill — fused conv + QKV split + L2norm (CANN aclnnCausalConv1dQkv)
  • chunk_gated_delta_rule — prefill delta-rule recurrence (CANN aclnnMegaChunkGdn)
  • fused_gdn_gating — decay gate computation (TileLang)
  • fused_sigmoid_gating_delta_rule_decode — fused decode recurrence (Triton-on-Ascend)

• Refactors ops_api.cpp NPU causal_conv1d_update implementation into a reusable tilelang wrapper (causal_conv1d_update_wrapper.cpp), shared by both the C++ model layer and the Python torch.ops dispatch.

• Introduces gdn_prefill_prepare as a unified interface for both CUDA and NPU prefill paths, encapsulating platform-specific fusion strategies.

• Enables qwen3_5 in model_platform_support.py for NPU.

Precision

• Layer 0 hidden_states: bit-exact match with C++ native path
• Layer 0 core_attn_out: max diff = 2.4e-4 (1 ULP in bf16)
• End-to-end output tokens: identical to C++ native for short prompts (temperature=0)

Test plan

• xllm serve --model_impl=python starts successfully with Qwen3.5-4B on NPU
• End-to-end inference produces correct, coherent output (no garbled text)
• Output token sequence matches C++ native path for short prompts
• Multiple prompts (Chinese/English, short/long) verified correct
• Run python_executor_test after fixing _EXPORTS sync
• Performance benchmarking vs C++ native path

[zhang-minchao]

为什么需要这个？

[yingxudeng]

格式检查没过

[yingxudeng]

是否会影响libtorch 组图的 qwen3.5