Pseudo-complex dot product and its gradient

[lamblin]

Before the rehaul of code generation and gradient that would be needed to fully support complex gradients in Theano, we can still wrap complex Ops as if operating on pairs of reals (more precisely, on arrays with an additional inner dimension of length 2).
For instance, this is what has been done for the RFFT's output and IRFFT's input in theano/tensor/fft.py.

The next bottleneck is the dot product, and there should be a way of expressing the gradients using {c,z}gemm. We should figure this out, and implement it.