Performance enhancement for inference

[nicolasnn]

On GPU hardware, in production scripts, post-processing takes twice as long as neural network inference.

On an NVidia GeForce GTX 1080 Ti (12GB RAM), for one full-size Sentinel-2 tile, CRGA OS2 UNet model, tile size 1024:

• With post-processing: ~6 minutes
• Without post-processing: ~2 minutes

A temporary workaround for speedup could be to skip post-processing by setting nodatavalues to None in inference.py. However, this would induce some artifacts, especially on images containing NoData.
A more long-term solution would be to include the post-processing inside the Keras model.

[remicres]

Starting from otbtf >=3.3.0 this would be easy to implement thanks to the otbtf.model.ModelBase.
We just have to implement the post-processing steps in the postprocess_outputs() method.