Calculate correct local cost, in order to decide on switching binary variables.
Separate out the temperature annealing into a separate neuron model. This allows to save CX state in the NEBM0-SA neuron, to have 16bit temperature resolution, and to save 1 ucode pass for all NEBM-SA neurons per time step.
Prevent conflicting variable switches. Initially, all neurons could switch simultaneously, which may substantially worsen the overall network state. This has previously been removed using refractive periods, which was merely a hack. The new neuron model allows, on average, only a single flip of variables that have conflicting cost terms.
Older neuron models will get removed with this PR.
Objective of pull request:
This PR resolves 3 issues in the QUBO solver:
Older neuron models will get removed with this PR.