[Feature] support rmsnorm_gated for qwen3.5

[wangna11BD]

Motivation

为支持 Qwen3.5 模型中带门控的RMSNorm（RMSNormGated）操作，新增了基于 Triton 的 fused RMSNormGated kernel 及对应的 RMSNormGated 层实现。

Modifications

1. 新增了基于 Triton 的 fused kernel rms_norm_gated_fwd_kernel

• 支持 swish/silu/sigmoid 三种 gate 激活函数
• 支持可选 bias、可选 z（gate 张量）

2. 新增了RMSNormGated 层实现

• CUDA + Triton 可用时，使用融合 kernel；其他平台（如 GCU）则降级到 PaddlePaddle 原生实现

3. 新增相关测试

Usage or Command

from fastdeploy.model_executor.layers.normalization import RMSNormGated

layer = RMSNormGated(hidden_size=2048, prefix="model.norm", activation="swish")

output = layer.forward(x, z=z)

Accuracy Tests

RMSNormGated 与 naive Python 实现的数值误差阈值设为1e-3

python tests/layers/test_rmsnorm_gated.py

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[Feature]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.

[paddle-bot]

Thanks for your contribution!