QKeras updates

[thesps]

New specific things:

• Add an Optimizer pass to set the rounding / saturation mode for the layer output. This is necessary to get decent with performance with the low precision quantized_relu layers. The pass is effectively disabled by default, since it may not be necessary for quantized_relu layers with higher precision.
• New softmax implementation. The softmax implementation was quite error prone, and contained several hard-coded values, so I have rewritten it. The new version also has an extra configurability: each LUT precision can be set independently. Although even without this configurability, the new version is more reliable.

[thesps]

The new softmax implementation needs to be propagated to the other converters. Since it is represented by a new layer class, we need to make sure the converter instantiates the right layer (the new Softmax, rather than `Activation).

[thesps]

I added the new softmax layer (with its two tables) to the other converters. I tested the onnx and TF models (successfully) but couldn't try to Pytorch one yet due to (I think) some issue with that model being from a much older Pytorch version.