makes it play nicer with the new rule that update values must have the same type
as their shared vars.

An alternative fix to several tests would have been to implement a
Tensor->Scalar cast, but I didn't do that because it's nice to just use Tensors.

Reason being there is a problem with the way we deal with strict.  Strictness
requires that a value has exactly the right type, but non-strictness implies
just about any kind of casting including in-exact casting.  What I would like
the default to be is something that means casting is ok as long as it is exact.

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.