had been upcast.  I did this for add, sub, div.  Mul looked tricky so I
didn't do it.

function.  This corresponds to what is already in theano.tensor.

when initializing a shared variable with a value of 0 because subsequent
floating-point assignments will be silently downcasted to 0.  At the same time,
it is slightly more annoying when it comes to float32 vs. float64 differences.

switched the optional parameter order of scalar_constructor so that it starts with (value, name, strict)

tensors.

still they are not perfect because there are cases where even the last two
dimensions could be broadcastable).

instead of 'Reshape' directly.  This is friendlier to subclasses.

single-precision floats.

compiledir is mapped to a tmpdir where an .so should go.