numexpr layer

[jacobpennington]

Similar to NumPy Layer class, it could be useful to have a NumExpr Layer class, along the lines of:

   ne_layer = NumericalExpression(parameters=['a', 'b', 'c'], expression='a*_x1 + b*_x2 + c*_x3')
   out = ne_layer.evaluate(input, another_input, state)

which would set fittable parameters named a, b, c under the hood and parse an arbitrary number of inputs into the incremented _x1, _x2, and _x3 variables. Benefit would be to easily implement simple models without needing to mess with building new Layer subclasses.

Details to figure out: specifying parameter shapes. Maybe parameters = {'a': (1,), 'b': (3,4), 'c': (1,2,3)} ?