Store quantized weights and calibration scales in the checkpoint.

[lukaszlew]

This is needed for inference performance as weight memory transfers are usually the limiting factor

[lukaszlew]

This is now fixed.