perf: avoid graph break for SiLUT when inferring (#4790)

This pull request simplifies and optimizes the implementation of the
`forward` method in the `ActivationFn` class within
`deepmd/pt/utils/utils.py`. The changes streamline the logic by removing
unnecessary condition checks and directly using `torch.where` for
computation.

I've evaluated this change using inference efficiency tasks from
LAMBench with DPA 3.1 3M model.

| System | Before: Avg Time ± Std (s) | After: Avg Time ± Std (s) |
Speedup | Success Rate |

|-------------------|---------------------------|--------------------------|---------|--------------|
| `catalysts_500.traj` | 211.82 ± 19.31 | **196.14 ± 18.11** | +7.1% |
100.0% |
| `inorganic_500.traj` | 204.62 ± 40.22 | **191.20 ± 36.44** | +6.4% |
100.0% |

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

## Summary by CodeRabbit

- **Refactor**
- Improved the internal logic of the SiLU activation function for more
streamlined processing. No changes to user-facing functionality.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Author: caic99

deepmd/pt/utils/utils.py

@@ -149,13 +149,10 @@ def silu_grad(x):
         self.const = float(silu(threshold))
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
-        silu_part = F.silu(x)
-        mask = x >= self.threshold
-        if torch.any(mask):
-            tanh_part = torch.tanh(self.slope * (x - self.threshold)) + self.const
-            return torch.where(x < self.threshold, silu_part, tanh_part)
-        else:
-            return silu_part
+        sig = torch.sigmoid(x)
+        silu = x * sig
+        tanh = torch.tanh(self.slope * (x - self.threshold)) + self.const
+        return torch.where(x >= self.threshold, tanh, silu)
 
 
 class ActivationFn(torch.nn.Module):