Fstd2nc

* More bug fixes for RSF support.
* Fixed auto-detect of legacy FST files in xarray.open_dataset.
This was broken in 0.20240401.1 after adding RSF detection.
* Fixed crash when using --interp and --turbo features at the same
time.
* Fixed --crop-to-smallest grid not working in some cases.
decodeGrid was sometimes changing ig1/ig2/ig3 identifiers and causing
a failure to find common grids.

* Fixed crash when importing fstd2nc.extra without having rpnpy or
librmn in the environment.
* Some fixes for RSF support. Detect more recent variants of header
structures (e.g. for librmn 20.1.0-alpha3 snapshot).
* Fixed missing auto-detect for RSF in xarray interface.
* Fixed address overflow issue with CCC file interface, which was
giving missing data for larger files.

* Some support for new RSF file container and fst24 data format.
* Fixed bug in fstpy bridge - nk no longer stored in Buffer.
* Fixed check for operational dictionary when opdict=True is used in
Python interface.
* Added _CRS attribute to xarray interface.
* Fixed bug when using dask distributed - the keys were not integers,
but rather numpy integers, which apparently was a problem.
* Some support for mask typvar '!'.
* Fix potential bug if used in future librmn release 20.1 - the
call signature for armn_compress now requires an extra 'swap_stream'
argument to control byte swapping of the decoded values.

* Fixed a bug in open_index() when the index contains pickled data.
* Fix some bugs in vertical bounds (when --bounds is used).
* Added some more fields to the list of accumulated quantities.
* Fix a bug when using .from_fstpy() followed by .to_xarray() on
a single record.

* Fixed bug in make_index / open_index when storing scalar parameters.
* Fixed bug in handling large files with masked fields. The mask
addresses were stored in 32-bits, causing an overflow in the address.

* Fixed preferred chunking when creating xarray objects via
open_dataset().
* Reduced the memory footprint of Buffer objects, and some small
speedups in the header decoding.
* Fixed CCC decoder. It was broken after a recent refactor, and there
was bug where the min value wasn't being added to the decoded
values.
* Added xarray backend for CCC files, so they can be opened with
xr.open_dataset(...,engine='ccc').
* Catch some possible runtime issues when fetching data from xarray
interface, such as if the file was removed from disk. Return
missing values if a problem occurs.
* Improved efficiency of --fstd-compat option (less padding used).
* Fixed a bug when using --bounds on surface data that has a
'sfc_type' dimension.
* Silence some warnings from recent vgrid libraries, when querying for
the existence of certain parameters.
* Added an 'accum' axis for accumulated quantities. Also added an
'--accum-vars' option to explicitly set fields that should be
treated as accumulated.
* Explicitly use a spherical globe in the projections returned by
fstd2nc.extra.get_crs().
* New helper methods for working with ridiculously large datasets
with xarray.
- fstd2nc.Buffer.make_index() pre-scans the input files and
generates an 'index' file on disk which contains all the variable
structures.
- fstd2nc.Buffer.open_index() reads the above 'index' file and
returns an xarray Dataset representing the FSTD data. This step
should be *much* faster since the files were already scanned
in the first step.