Make learning snip more efficient

[tbekolay]

Currently the communication protocol introduced in #89 is not too efficient. We store metadata in the userData struct, and read it on every timestep to determine what part of the userData struct is used for each learned connection.

It would be more efficient to make nengo_learn.c a jinja template like nengo_io.c so that the metadata and slices of the userData struct associated with each learning connection is known a priori and don't need to be read on every timestep.

If we have time to really bust out the profiler, we should also see if there is any benefit to doing all of the IO in nengo_io.c rather than making separate channels and also doing IO in nengo_learn.c.