Adding more loss options during training

[gattia]

The first commit is just a documentation error I noted.

The main changes are thoroughly described in the second commit. Briefly:

• I added the softmax option for building the 'avg_dice_no_reduce'.
• I also added new functionality that enables calling multiple additional losses during training.
• To go along with the multiple loss, I added the ability to use DiceLoss to return the loss of just a single class/tissue, this way we can monitor how individual tissues are progressing during the training process.

This is probably still a bit of a work in progress - but I wanted to see if there are any thought or feedback as I work on this.

I've added the second commit message w/ the majority of the changes below to make the content easier to see in this pull request:

• config.py updated to take a list called LOSS_METRICS.
  This is a list of extra loss metrics to run during train/val steps.
  Each item in the list is structured as:
  [[loss_type, activation], weights]. The first item is a list that
  mimics the current cfg.LOSS input into build_loss
  and the second item is the weights.

• losses.py now explicitly includes "softmax" version of
  "avg_dice_no_reduce" ("avg_dice_no_reduce", "softmax")
  which calls DiceLoss. Though, this might be redundant,
  becuase it does the exact same thing as calling
  ("avg_dice_no_reduce", "sigmoid"). From this standpoint,
  it might be useful to break LOSS into the actual loss part
  ('avg_dice_no_reduce') and the activation
  ('softmax'/'sigmoid').

  The current implementation was a bit confusing to me -
  I thought I had to pass the softmax as the activation to
  DiceLoss which caused issues. Breaking it up into
  these parts would be clearer.

• losses.py got a new function that creates a one-hot-encoded
  set of weights if an integer is inputted instead of
  a list of weights. This is useful it the goal is to just find the
  loss of a single tissue/class during training.

• build_loss was updated to enable building the additional
  loss functions mentioned in config.py above.

• trainer.py was updated so that it builds a list of the loss metrics
  mentioned in config.py and enabled by the updates to
  build_loss.

• reduce_tensor in utils.py was updated so that if the reduce was
  'none' and the weights were one-hot encoded, it scales the weights
  value so that it makes sense. Otherwise, when keras does its
  built in reduce it averages over all of the zero dims making the loss
  seem artificially low (by a factor of the number of categories/classes
  in the loss).