mrmatrix: coarsen should create nan_values at each resolution?

@pkerpedjiev : Am I understanding it correctly, that there should be a nan_values at each resolution? It is not producing that right now. Without that, it would still need to consult the large matrix at resolution 1 to calculate averages. Assign back to me if there's work to do.

`ts_to_mrmatrix_test.py`:
```
class ParseTest(unittest.TestCase):
def test_parse(self):
...
self.assertEqual(list(hdf5['resolutions']['1'].keys()), ['nan_values', 'values'])
...
self.assertEqual(list(hdf5['resolutions']['2'].keys()), ['values'])
```