PyTorch model to CoreML model and corresponding kCVPixelFormatType of CVPixelBuffer

## ❓Question

I use coremltools to convert a PyTorch model to CoreML model and `kCVPixelFormatType` of `CVPixelBuffer` the model should receive confused me.
There are two scenarios:
1. Use the default `color_layout = ct.colorlayout.RGB` to convert the model. This CoreML model receives `CVPixelBuffer` with `kCVPixelFormatType_32ARGB` and works well. Besides, it outputs wrong answers when the `CVPixelBuffer`'s `kCVPixelFormatType` is `kCVPixelFormatType_32BGRA`.
2. Use `color_layout = ct.colorlayout.BGR` to convert the model. The CoreML model still outputs different and wrong answers when the input `CVPixelBuffer`'s
`kCVPixelFormatType` is `kCVPixelFormatType_32BGRA`.

I wonder how to convert the model and enable the CoreML model to correctly work on `CVPixelBuffer` with `kCVPixelFormatType_32BGRA` because it may take lots of steps to get `kCVPixelFormatType_32ARGB` format `CVPixelBuffer`.

Thank you!

Start by reproducing the two PyTorch-to-CoreML conversions described in the issue, using both RGB and BGR color layouts and the two CVPixelBuffer pixel formats. Compare the model outputs for kCVPixelFormatType_32ARGB and kCVPixelFormatType_32BGRA. Done means the correct conversion or supported input behavior for BGRA is established and documented.

Written by the indexing model from the issue text.