Dimarray

The main point of the release is to support python3.

Along this way, the dimarray.geo package was removed, as will be all non-essential components in future releases.
For geo-physical applications, use xarray.

In previous releases, a fix to cartopy classes was done by modifying cartopy.crs module.
In this release, the fix only applies to dimarray.compat.cartopy_crs, as it should.

Use scipy.interpolate.RegularGridInterpolator instead of basemap interp.
This also involves new parameters for `dimarray.geo.transform`, `dimarray.geo.transform_vectors` and `dimarray.interp2d` , accordingly to the new underlying function.
In particular this may require the use of keyword argument `bounds_error = False` for new axes that are out-of-bound with respect to original axes.

- deprecated / removed:
- `aggregate` (not useful)
- `DimArray.groupby` (not really a multi-dimensional operation, alternative: dima.to_pandas().groupby(...) )
- modified behavior
- multi-dimension indexing : now "orthogonal" by default. Use `DimArray.take(..., broadcast=True)` to broadcast arrays on the same axis, consistently with numpy.
- renamed:
- `DimArray.group` => `DimArray.flatten`
- `DimArray.ungroup` => `DimArray.unflatten`
- `GroupedAxis` => `MultiAxis`
- `align_axes` => `align`
- new features:
- `align` : new `join=` parameter "inner" or "outer"
- slicing on ordered numerical axis : a[2.:4.] on [1.5, 2.5, 3.5, 4.5] now works and return values on [2.5, 3.5]
- `open_nc` to return a `DatasetOnDisk` : allow examination of datasets without actually reading it (somewhere between read_nc and the original netCDF4 module)
- `DimArray.iloc`, `DimArray.loc`, `DimArray.sel`, `DimArray.isel`: convergence toward xray/pandas API
- `DimArray.nloc` : like DimArray.loc but nearest-neighbor indexing
- numpy.datetime64 can now be written to netCDF, and converted back (conversion adapted from on `xray.conventions` functions)
- internal:
- indexing rewritten from the bottom up (basically: numpy.argsort + searchsorted and subsequent check, when more than one element is searched, `nonzero` otherwise)
==> much shorter source code
==> more efficient calculations (reindexing-speed comparable with xray, a factor 2 slower than pandas)
==> no dependency on pandas anymore
- netCDF I/O rewritten, too, now relies on DatasetOnDisk, DimArrayOnDisk, AxisOnDisk, AxesOnDisk, AttrsOnDisk classes.
- AbstractDimArray, AbstractDataset and AbstractAxis classes from which DimArray and DimArrayOnDisk derive
==> indexing and formatting unified
- experimental:
- dimarray.geo.crs.get_crs() : returns an instance of a cartopy CRS class based on a dictionary of CF conventions
- dimarray.geo.transform() : transform and interpolate a dimarray from one coordinate system to another
- dimarray.geo.transform_array() : transform and interpolate two dimarrays

Temporary roadmap for future release(s)

Mostly, simplify and thin down the source code, make it more readable:
- dataset : do not make it an instance of OrderedDict with management of axes copy/refs as it is now, but make it a new object.
- simplify indexing in abstract classes (keep only orthogonal indexing, provide array index broadcasting as a separate function)
- simplify axes definition (allow only a few possible methods, to reduce the amount of checks)
- deprecate the dimarray.geo.GeoArray object, and flatten the sub-package for the user
(everything accessible from dimarray.geo (not crs etc...))
- review the library of functions provided, only keep the most useful ones
- simplify the cascade of classes in the netCDF I/O part
- remove the global option file?
- improve unit tests to have less redundancy, more readability

This release is not a true release, but an intermediate milestone which contains a number of bug fixes and improvements, as well as a beta version of the geo module. It will be followed soon by a more significant update (probably 0.2 version), with improved indexing, netCDF I/O, metadata handling and Dataset functionality. These changes might introduce minor back-compatibility issues, and thus the 0.1.9 release is more useful for those who would have problems using the 0.2 version.