Float 128 usually not saved correctly

[matthew-brett]

The nifti standard allows us to save float128, but numpy float128 is almost invariably not what nifti imagines to be float128 - it's usually intel 80 bit padded to 128 bits.

So, we'll save a confusing file which will load fine in numpy but be completely wrong elsewhere. So we should probably raise an error for float128 or do so when we're not running on an IBM S/360 or whatever has a real float128