Fix for memory issue

[s7yoewer]

Memory issue is partly fixed (#41). The memory now rises much slower than before. Observation structures are deallocated at the end of each assimilation time step. For global arrays, OMI internal function is called, local arrays are deallocated manually. The functions are called in prepoststep with a flag to make sure they are called during the second call of prepoststep. Further changes:

• observation file is closed after loading --> contributed to memory issue
• minor bug fixes --> did not really had an effect on the current simulations bur are still fixed

[s7yoewer]

There is still some memory problems somewhere. Per year (with daily observations assimilated), the max memory is still rising by ca. 1.7GB (llview). However, this means I can conduct my long simulations. Before, it was something around 10GB per year. I am not sure where this is now coming from but the fix is enough for me for now.

[s7yoewer]

But as far as I can see this is also the case when I do not use OMI, maybe half of the memory rise compared to OMI because two observation types are processed for OMI. I will keep my changes and not investigate further.