Memory usage for DICOM files

[parneshr]

While not a problem for the majority of everyday use, for certain memory intensive operations such as converting coil data to nifti (DICOM dataset for these are ~ 5Gb), the memory use of nibabel blows up considerably. I have isolated this to the _scale_data functions in the nicom/dicomwrappers.py file. By default the Rescale slope and intercept are converted to python floats which are 64 bits. Hence data is cast to 64 bit floats in _apply_scale_offset. Perhaps numpy.float32 should be used as a default instead?