load_data, precision for bad_flag

[castelao]

The load_data() loads the raw data into a masked array. It uses the bad_flag to identify the non valid data to be masked. The default tolerance for the masked_value is 1e-8, but it's common to use values as low as 1e-29 on the bad_flag.

I temporary set atol to 1e-30, but this must be improved. Probably the best solution is to drive the loading process with numpy.fromstring()