You don’t have to use the built-in mxnet imread type functions for you reading in the image. For example, you can write your own custom DataLoader/Dataset that uses the jpeg decompression library you want on your images. Once you have read in the image with your library, then you can create an ndarray of the image pixels and use that in your training loop.
It depends which MXNet you’re using and where you got it from. Just taking mxnet-cu92 from pip as an example, it does appear to be using LIBJPEG_TURBO. See here. And when compiling MXNet from source, you can configure as you wish.