Support writing to FITS

[sjperkins]

As dask processes datasets in chunks, It would be useful to support writing to FITS files in chunks. The alternative is aggregating all chunks into a single large array before writing to disk, which is untenable with the data sizes that we currently encounter.

In the reading case, each chunk is handled as follows:

xarray-fits/xarrayfits/fits.py

Lines 66 to 72 in 3360d94

	def _get_data_function(fits_proxy, h, i, dt):
	if fits_proxy.is_memory_mapped:
	data = fits_proxy.hdu_list[h].data[i]
	else:
	data = fits_proxy.hdu_list[h].section[i]
	return data.astype(dt.newbyteorder("="))

Either the data or section attributes are accessed, depending whether the file is memory mapped on a local filesystem, or remotely accessed. Presumably these attributes can be used to support chunk writes.

One concern I have is whether writing from multiple threads/processes will be handled properly. This is probably OK in the remote case, and I expect that the OS will handle paging the writes between memory and disk in the memory mapped case.

/cc @bennahugo @o-smirnov