Harmonize Theano rfft with numpy rfftn

[slefrancois]

Theano's fft functions correspond to numpy's fftn, for n-dimensionnal transform. This is confusing because numpy.fft.fft is the 1-D transform.

For both tensor.fft and and gpuarray.fft, we should rename the user functions (rfft > rfftn, irfft > irfftn) to fit numpy. The current user functions should be deprecated with a warning. The GPU replacement optimizers must be changed too.