Parallel FFT objects should be aware of the global shape of the output

[sbrisard]

In module janus.fft.parallel transform objects are aware of the global shape of the input (the shape attribute), the local shape of the input (the rshape attribute), the local shape of the output (the cshape attribute), but not the global shape of the output.

This leads to rather clumsy constructs when one wants to gather the results of a FFT transform. The array which holds the global output must be allocated as follows

if comm.rank == root:
    y = np.empty((transform.shape[0],) + transform.cshape[1:],
                 dtype=np.float64)
else:
    y = None
comm.Gatherv(y_loc, y, root)

It would be much nicer if transform had an attribute holding the global shape of the output.

[sbrisard]

In 289062e117713dc0ce61db48edaa5377ebe51c97, FFT objects define a global_oshape attribute (tuple).

[sbrisard]

In d56debd911a6d2eaa56234ad1196488c828e9eca, renamed attribute of FFT objects: shape -> global_ishape (see Issue #7).