Fstd2nc

* New --rpnstd-metadata-list to select a subset of RPN metadata
attributes to copy to the netCDF file.
* Made --minimal-metadata the default. The old behaviour (copying all
RPN metadata) can be triggered with a new --rpnstd-metadata option.
* Coordinate-related RPN metadata is no longer copied, since the values
are already encoded in the coordinate variable.
* Restored support for 'A', 'B', and 'G' grids.
* Fixed the repeated longitude in global grids - value is now 360
instead of wrapping back to 0.
* Use consistent lat/lon/station coordinates for series data when using
a --filter on the ip3 values.
* Unused axes are now pruned from the file. Mostly affected series
files, which had an extra time axis from the unused datev values.
* Vertical axes are now encoded as floats, since that's the precision
of the original ip1 values. The pandas version of the iterator was
upconverting these to doubles, which would cause different results
depending on whether pandas was installed or not.

* Improvements to xarray interface to utilize the CF conventions
metadata (such as auxiliary coordinates).
* Add station, leadtime, and reftime as auxiliary coordinates in series
files.
* Use the date of validity by default. Old behaviour (date of origin)
can be restored by using --dateo or --forecast-axis options.
* Added 'leadtime' and 'reftime' auxiliary coordinates to encode the
forecast information.
* Progress bar is enabled by default. Can be turned off with
--no-progress option.
* Added CF metadata for various grid projections (rotated pole, polar
stereographic, lat-lon).
* Decode surface types (ice, land, sea, etc.).
* Decode depth coordinate for ocean model output.
* Fix decoding of masked data in raw binary (X32) format.

* Fixed unlimited dimension specification for PyDAP interface.
* Workaround for bug in numpy 1.14 when using masked structured arrays.
* Fix forecast times for series files.
* Don't discard LA/LO fields. They needed in some circumstances such
as trajectory files.

* Faster scanning of input file headers.
* New --progress option to show a progress bar for long conversions.
* More speed optimizations for RPN date and IP1 value conversions.
* Faster determination of dimensions from the record headers, using
pandas package.
* Restructured the package so each feature is in a separate Python
module, for easier development.
* Added xarray interface to facilitate custom netCDF conversions from
Python.

* Faster I/O when converting multiple input files.

* Support for ''scale_factor'' and ''offset'' attributes in metadata
files, as a rudimentary form of unit conversion.
* New --zlib option for compression of netCDF files.
* Some speed optimizations for converting to netCDF4:
- Chunk size is set equal to FSTD record size.
- FSTD records are read in order, to better utilize buffering /
read-ahead on the input file(s).