fix release of pytorch 1.0, semantic segmentation, and instance segmentation branches

[shelhamer]

I somehow packaged up an incomplete branch for release that most importantly (1) was not fully compatible with pytorch 1.0 and (2) was missing our model with an unshared metric for positives and negatives. The necessary fixes and grooming to correct that are bundled together here.

• fix train + evaluate for pytorch 1.0
• fix collate for pytorch 1.0
• fix model override for cofeat-late-lite
• fix dataset dict: trivial syntax error
• model: include unshared metric with separate positive/negative weights
  (which we found to be more effective for semantic segmentation)
• model: switch to nn.init helpers and new mutable module interface
  for surgery
• data: notebook for few-shot datasets
• data: include slug in aux
• rename DIOS -> interactive: interactive is more descriptive, and only
  the early fusion models are closely related to DIOS
• drop stale loss module: pytorch 1.0 handles 2d inputs
• drop unused data arg multi
• drop stale usage of autograd.Variable

close #3