In the MTG FCI FDHSI L1C reader, get the projection coordinates and
extent information from file.  For the projection coordinates, get those
from /data/mtg_geos_projection rather than from /state/processor, as
is consistent with the documentation.  The extent information was
previously hardcoded but now obtained from data/channel/measured/x and
data/channel/measured/y, which contain this information in radians.
This commit fixes pytroll#840 although there are still some issues with the
resulting image that need to be resolved.

In FCI projection area extent, use the pixel outer corners rather than the centres.

Newly the FCI reader is using new fields it wasn't using before.
Make sure the test routines are mocking these fields so that the tests
can run.

Since we still support Python 2, replace f-string by regular string in
FCI L1C FDHSI reader

Remove the unused ``_fallback_area_def`` attribute from the FCI L1C
FDHSI FileHandler class.

Added to the reader module docstring information on how the geolocation
is performed.

Earlier this year, I programmed a workaround for a bug in xarray.  This
bug got fixed in xarray 0.13.  This commit disabled the workaround for
sufficiently recent versions of xarray.  Hopefully this will improve
performance.

The FCI reader now uses the h5netcdf backend.  See the comments on issue
972 for the rationale.  The netcdf4 backend has a memory leak and even
worse performance than the h5netcdf backend for opening and closing
files many times, which the filehandler does.

This addresses pytroll#972 although I'm not yet fully satisfied.

When reading attributes, do not access the data variable.
Accessing the data variable will trigger netcdf_utils to
actually open (and close) the netcdf file using xarray,
causing unneccessary overhead.  Instead, use the fact that
the netcdf reader has cached this information.

More caching may be needed.

This commit is related to pytroll#972.

For channels with the same resolution, don't recalculate the area extent
as this is a performance hit.  Instead, cache the area extent between
channels with the same resolution.

Fix a bug in the FCI performance optimisation, when caching area extents
do not share this cache between different instances.

FCI filehandler was setting nlines and ncols as an attribute, but this is
channel dependent.  Store as a dictionary instead.  I don't know why
this did not trigger any problems before the recent optimisations.

Bugfix in fci reader, now that I'm sometimes returning stuff from cache,
make sure the side-effects such as setting other attributes has been
carried out already...

Needlessly storing the state is just asking for problems, as I
experienced.  Don't store ncols and nrows as attributes, instead pass
them on between methods.  This is not a property of a filehandler.

In the FCI reader, do not store the area definition as a first-level
instance attribute.  This is different between channels.

Instead of caching the area extent, cache the full area definition.

Do not duplicate the performance improvements from the other branch.

In the FCI reader area definition, pass the inverse flattening, as
included in the file, from the reader to the `AreaDefinition` into
pyresample.

The scale factors and add offsets for the MTG FCI test data as published
in 2019-09 are factually incorrect.  This probably causes the
differences between the pytroll and FCI geolocations.  Until this is
fixed from EUMETSATs side, hardcode corrected add_offset and
scale_factor values.

Also correct for an off-by-one error in the scale factors.

The scale factors and add offsets for the MTG FCI test data as published
in 2019-09 are factually incorrect.  This probably causes the
differences between the pytroll and FCI geolocations.  Until this is
fixed from EUMETSATs side, hardcode corrected add_offset and
scale_factor values.

Similarly, there was a small error in the semi minor axis.

Also correct for an off-by-one error in the scale factors.

The FCI semi-minor axis correction does not improve anything, remove
this hardcoding correction.

… fix-fci-projection

Expand the mocking of variable attributes for the routine testing the
FCI reader.  One attribute was missing, and none were created in the
"/attr/..." caching in which they are accessed by the reader.

Remove the hardcoded scale factor, add offset, and sweep.  Those had
been added to correct for errors in the 2019-09 release of FCI testdata,
but should be fixed in the upcoming release, so should be no longer
needed.

From the FCI reader, remove an unused hardcoded dictionary containing
extent information.  This information was only used for debugging, but
is redundant relative to what is available in the file, assuming a full
disk image is square.

In the FCI reader, make minor fixes to comments and docstring, and
replace a RuntimeError by a more specific ValueError if an invalid
calibration key is passed.

Apparently, PEP 257 insists that the closing quotes for a single-line docstring must be on the same line as the rest of the docstring.  I will abide by what is apparently the community standard, although I will hereby note my (minority?) preference for the "closing quotes on next line" alternative.

Co-Authored-By: Martin Raspaud <martin.raspaud@smhi.se>

Remove an empty line immediately after a method docstring, in order to
satisfy the flake8-docstrings plugin.