More realistic Ca2+ based learning rules: Urakubo, Kinase

[rcoreilly]

start by trying to make a simple replication of core elements of Urakubo et al? Look up recent work in this area to see what else is out there, e.g., bcm guy..

[rcoreilly]

see examples/urakubo and ccnlab/kinase repo -- now integrating.

[rcoreilly]

after extensive work on this, the kinase family of learning rules covers a lot of biological ground and works well.. now moving these into kinasex exploratory Prjn types instead of having in the main axon Prjn, which will just retain the SynSpkTheta and NeurSpkTheta algorithm variants (just as a Neur flag with Syn being the default).

SynSpkTheta works significantly better than Neur on LVis model, by computing Ca values at the synapse level, which captures more detailed temporal correlations among send, recv neurons than just using separate neuron-level variables. The extra computational cost of synapse-level updating is mitigated by optimizations.

The kinasex contains SynSpkCont and SynNMDACont variants that do continuous learning on CaP vs. CaD with peak-finding -- no local dependency on explicit theta cycle. NMDA version closely approximates Urakubo detailed model and works. Both are significantly more computationally expensive and do not present any obvious computational benefits..