NXTomoReader is too slow when using angular subsampling

The problem is described in #335

from nabu.io.reader import NXTomoReader

nx2 = NXTomoReader(fname, sub_region=(slice(None, None, None), slice(None), slice(None)))
%time projs = nx2.load_data()
projs.data.nbytes/1e9/2/47.5 # 1.4 GB/s

nx2b = NXTomoReader(fname, sub_region=(slice(None, None, 2), slice(None), slice(None)))
projs.data.nbytes/1e9/2/(10*60+41) # 103 MB/s !!

For now the only solution is to read all the data and do the subsampling in memory.