Add FoldAxes transformer

[cboulay]

If we have a 3+ dimensional tensor (e.g., time, channel, frequency) that we want to pass through a simple linear transform (e.g., affine, or some models from ezmsg.learn) then we may want to fold channel and frequency together into a single dimension. The new axis will probably be a coordinate axis named something like "ch-freq" and its .data entries will have some pattern constructed from the combination of the two input axes. e.g.: "1,2.0", "2,2.0", "3,2.0", ..., "62,100.0", "63,100.0", "64,100.0".