Alpha initialization must match weights and activations maximums

[frneer]

Alphas need to be initialized to the max of the weights → Post pre-training
Maybe set it as Quantizers default?
Activations need to be initialized to the max of the activations, but… they aren’t store

[frneer]

main...colo/model_evalution#diff-68e0a4ab79155ef94564c4be31a66dd9ec8c7604a1eb14cad3a821c6c52fd9deR472