fix(dpmodel): fix normalize scale of initial parameters (#4774)

The current scale is too large. This PR makes it consistent with PT.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

- **Refactor**
- Improved the initialization of certain neural network parameters for
enhanced stability and consistency. No changes to user-facing
functionality.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Signed-off-by: Jinzhe Zeng <jinzhe.zeng@ustc.edu.cn>
Signed-off-by: Jinzhe Zeng <njzjz@qq.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Author: njzjz

deepmd/dpmodel/utils/network.py

@@ -105,9 +105,18 @@ def __init__(
         # only use_timestep when skip connection is established.
         use_timestep = use_timestep and (num_out == num_in or num_out == num_in * 2)
         rng = np.random.default_rng(seed)
-        self.w = rng.normal(size=(num_in, num_out)).astype(prec)
-        self.b = rng.normal(size=(num_out,)).astype(prec) if bias else None
-        self.idt = rng.normal(size=(num_out,)).astype(prec) if use_timestep else None
+        scale_factor = 1.0 / np.sqrt(num_out + num_in)
+        self.w = rng.normal(size=(num_in, num_out), scale=scale_factor).astype(prec)
+        self.b = (
+            rng.normal(size=(num_out,), scale=scale_factor).astype(prec)
+            if bias
+            else None
+        )
+        self.idt = (
+            rng.normal(size=(num_out,), scale=scale_factor).astype(prec)
+            if use_timestep
+            else None
+        )
         self.activation_function = (
             activation_function if activation_function is not None else "none"
         )