Further improve appearance of color/colorbar for non-binary segmentations

[fedorov]

Follow up on #287

I thought more about this after realizing that DICOM standard allows for specifying color for fractional/probabilistic segmentations.

I would suggest we discuss the following as an idea to further improve appearance of those segmentations.

1. each distinct structure (as defined by category/type combination) is assigned a distinct color (same as we do for binary seg)
2. instead of fading from dark to brigher color, we instead go from 100% to 0% transparency for the given color
3. to handle color bars, we create a separate colorbar oriented horizontally in the corner, and show the label above each colorbar

[fedorov]

Key points were discussed in this thread: #245 (comment).

We should meet again to brainstorm how/if this should be addressed.

[CPBridge]

I think my concerns about the current behavior are summed up in this comment.

I had suggested at one point gradually fading the colormap towards transparent as we get towards zero, rather than abruptly switching from opaque to transparent going from 1 to 0. I have some sympathy with the idea that the value should control only color and not transparency, but in my opinion we are already going against that because a value of 0 makes the segmentation transparent and 1-255 make it opaque. So fading doesn't seem any worse than the current behavior in this regard.

One alternative idea could be to keep the current behavior of an abrupt transition but allow the user to use a slider to control the transition point between between transparent and opaque, rather than hardcode it at the 0-1 boundary. And maybe use a different default value of half the range (since this gives the position of equal likelihood of positive and negative class)