DICOM `scale=preserve` not working as intended and performance consideration

[jpambrun]

The implementation of the DICOM image decoder has some limitation and would benefit from some tweaks.

1. It's using DCMTK's getOutputData() which renders presentation pixels rather than returning modality values. For CT image, modality values follows a specific scale of Hounsfield unit where water density shows as 0 and air density is represented by -1000. getOutputData() looses this information. getInterData() should be used. Otherwise the scale=preserve option doesn't really work as intended.
2. getOutputData() is configured to output uin64 which seems inefficient and wasteful. For most modality the dynamic range is 12 bits at most. I don't know of any device or technology that would produce 64 bits of dynamic range. Even if it did, it's probably best to make this dynamic rather than hard coded.
3. Looking at the implementation, it seems it does a string comparison for the scale parameter for every single pixel. This should be taken out of the critical loop.

I have made most of the proposed change here some time ago. The scale implementation is missing, but it should be easy to add (and more efficient?) with by scaling the output between tf.reduce_min and tf.reduce_max.