Implement fragment writer which proceeds one column at a time

[kounelisagis]

This PR implements a new fragment writer class that writes one field (column) at a time, rather than writing all fields together in cell batches. This enables writing fragments where each dimension and attribute is processed independently - the caller opens a field, writes all of its tiles sequentially, closes it, and then moves to the next field.

The writer accepts pre-filtered WriterTileTuple objects and handles both dense and sparse arrays. For dense arrays, the tile count is computed automatically from the domain; sparse arrays require explicit MBRs at finalization.

Closes CORE-462

---

TYPE: FEATURE
DESC: Implement fragment writer which proceeds one column at a time

[rroelke]

The MBRs piece I hadn't really considered before this. I suppose we'll probably compute it in the execution node which filters the coordinate tile data. It's not really obvious because of multi-dimensional arrays, of course. Let's leave it this way for now.

EDIT: I'll suggest instead that you have a separate method to attach the MBRs. Consolidation will surely process the dimensions first, and so it will be able to compute the number of tiles and attach MBRs to this writer potentially before it begins processing the attributes. We can attach the MBRs and then free the intermediate memory to re-use for attribute processing. Whereas requiring them for finalize means we will have to keep them in memory throughout all of the attribute processing.