Inconsistent treatment of fill_value and missing_value during cubewrite

[blimlim]

This issue is related to cubewrite refactoring work in #99, where we are extracting the missing_value/fill value modifications into a separate function.

The code is not currently consistent in the way it treats the fill value's type. It first sets a fill value depending on the type of the cube's data, and then writes this to the cube's missing_data attribute:

um2nc-standalone/umpost/um2netcdf.py

Lines 331 to 339 in c1a4a54

	# Set the missing_value attribute. Use an array to force the type to match
	# the data type
	if cube.data.dtype.kind == 'f':
	fill_value = 1.e20
	else:
	# Use netCDF defaults
	fill_value = default_fillvals['%s%1d' % (cube.data.dtype.kind, cube.data.dtype.itemsize)]
	cube.attributes['missing_value'] = np.array([fill_value], cube.data.dtype)

where the missing_data attribute is forced to match the cube data's type.

When the fill_value is used later as an argument to to the sman.write function, the same type conversion is not applied, e.g:

um2nc-standalone/umpost/um2netcdf.py

Lines 391 to 395 in c1a4a54

	sman.write(cube,
	zlib=True,
	complevel=compression,
	unlimited_dimensions=['time'],
	fill_value=fill_value)

This means there can be a mismatch between the missing_value and fill_values types. E.g. if the cube.data.dtype is np.float32, then the missing_value will have type float32, while the fill_value supplied to sman.write will use the default python float64 type.

@MartinDix do you know if there are reasons that the type needs to be treated differently in these two places? In terms of turning this part of the code into a separate function and adding unit tests, I think it becomes simpler if we are able to convert both the missing_value and fill_value to match the cube.data.dtype.