Pack Neural Propagation Instructions tightly to spare space on GPU

[davids91]

Refer to Neurons.md

• set a reasonable maximum expected size for each propagation step ( e.g. what is the most number of supported weights )
• Document the limitations thoroughly
• Modify Propagation instructions to be a vector of contiguous byte arrays instead a 2D vector
• implement compression of each item during propagation instruction generation
• implement De-compression of each item during neural propagation in the kernels