Write a custom CUDA kernel to optimize `SwAT` (Swish-Arctan) activation function.

Formula: f(x) = x * sigmoid(arctan(x))

Problem Analysis:
1. Computationally Intensive & Memory Bound: The operation is element-wise but involves a chain of two transcendental functions (arctan, sigmoid contains exp).
2. Operator Chaining: A standard PyTorch implementation creates intermediate tensors for arctan and sigmoid, wasting memory bandwidth.

Optimization Strategy: Fused Element-wise Kernel with Vectorization

1. One-Thread-per-Element: Map each element to a CUDA thread.

2. Vectorized Loads (float4): Use `float4` to process 128 bits per memory transaction.

3. Fused In-Register Math:
   - For each element `x`:
     `atan_val = atanf(x)`
     `sig_val = 1.0f / (1.0f + __expf(-atan_val))`
     `result = x * sig_val`
   - All computations are fused in registers.

4. One-Pass: Fuse all steps into a single read-compute-write kernel.
  
Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is:   
  
```python
import torch
import torch.nn as nn

BATCH_SIZE = 4096
HIDDEN_DIM = 4096
SHAPE = (BATCH_SIZE, HIDDEN_DIM)

class SwAT(nn.Module):
    """
    SwAT Activation.
    Revisiting activation functions: empirical evaluation for image understanding and classification
    https://link.springer.com/article/10.1007/s11042-023-16159-2
    Formula: f(x) = x * sigmoid(arctan(x))
    """
    def __init__(self):
        super(SwAT, self).__init__()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return x * torch.sigmoid(torch.arctan(x))

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.act = SwAT()
    
    def forward(self, x):
        return self.act(x)

def get_inputs():
    input_tensor = torch.randn(SHAPE, dtype=torch.float32) * 5.0
    return [input_tensor.contiguous()]

def get_init_inputs():
    return []