buffer writediagfi outputs?

[emillour]

writediagfi could become a bottleneck when in parallel as all data is sent to master which then writes.

However, even in serial and in 1D writediagfi slows down the code quite a bit because it opens, fills and closes the output files every time it is called.
A possible optimization would be to buffer all the outputs to be made throughout the physics time step and only open, write and close once per time step.

(NB: using XIOS for the outputs also alleviates the problem as it buffers its outputs efficiently)