Refactor BufferWrapper

[sk1p]

Right now, BufferWrapper is mixing a few concerns. I think it would make sense to break it up a bit:

• Buffer declaration: basically UDF.buffer
• "Partition-shaped" result buffer: has a partition slice set, might have a roi etc.
• "Dataset-shaped" result buffers: contains the result "for the whole dataset" (which might be partial results still, basically what comes out of merge)

On another axis, it might make sense to make another split on the kind axis, as there are often special cases for nav and maybe sig, compared to single. If we have a PartResultBuffer, and inheriting from that, a NavPartResultBuffer, this should clear things up a bit.

The reason would be to have more static guarantees and with that also a better IDE experience, and to make the code easier to understand and modify. It should also make the interfaces exposed to users easier to understand - right now, the user just gets a BufferWrapper, but I can imagine that from the docs alone it's not clear how that is used in all the different cases.