SeparableConv2d implementation

I'm trying to implement the Xception model based on tensorflow Keras implementation (and also comparing to the pytorch version). Here is my implementation of the SeparableConv2d in candle:

```rust
use candle_core::{Error as CandleError, Tensor};
use candle_nn::{conv2d_no_bias, Conv2d, Conv2dConfig, Module, VarBuilder};

pub struct SeparableConv2d {
depthwise: Conv2d,
pointwise: Conv2d,
}

impl SeparableConv2d {
pub fn new(
in_channels: usize,
out_channels: usize,
kernel_size: usize,
stride: usize,
padding: usize,
dilation: usize,
vb: VarBuilder,
) -> Result {
let depthwise = conv2d_no_bias(
in_channels,
in_channels,
kernel_size,
Conv2dConfig {
stride,
padding,
dilation,
groups: in_channels,
},
vb.pp("depthwise"),
)?;

let pointwise = conv2d_no_bias(
in_channels,
out_channels,
1,
Conv2dConfig {
stride: 1,
padding: 0,
dilation: 1,
groups: 1,
},
vb.pp("pointwise"),
)?;

Ok(Self {
depthwise,
pointwise,
})
}
}

impl Module for SeparableConv2d {
fn forward(&self, xs: &Tensor) -> candle_core::Result {
xs.apply(&self.depthwise)?.apply(&self.pointwise)
}
}
```

But when I compare the output tensors of the original model and the candle version (with the same pre-trained weights, transposed in channel_first format for candle) using patdiff, I notice a high deviation from the candle sepconv output. Could someone confirm the SeparableConv2d implementation is correct ?

Start with the posted SeparableConv2d implementation and compare its depthwise and pointwise convolution configuration with the TensorFlow Keras and PyTorch implementations, using the same transposed channel-first weights. Confirm whether the outputs match; done means identifying whether the deviation comes from the convolution implementation or weight and layout handling.

Written by the indexing model from the issue text.