Default fill value dtype does not match that of data for masked arrays

[jakirkham]

The fill value type chosen for integers is always int64 and always 999999 even when that doesn't make sense numpy.ma.MaskedArray.fill_value. For instance, take the type uint8. In this case, the fill value is way outside of bounds and ends up getting converted to 63. If the behavior must be preserved, then for uint8 the fill value should be 63 and it's dtype should be uint8. Alternatively, better fill values could be picked that work on a range of types (i.e. 0 for integers and nan for floats and complex) or something similar.