Introducing Typed Ops using Scala Types instead of TensorFlow Types

[eaplatanios]

Ops and outputs are now also typed (similar to tensors). Their type parameters are the Scala equivalent of TensorFlow types. For example:

// These are valid types:
Tensor[Int]
Output[Int]
Output[Double]
Op[(Output[Int], Output[Double]), Output[Long]]

// while these are not:
Tensor[INT32]
Output[INT32]
Output[FLOAT64]

This also results in:

• Simplification of many interfaces. For example, Dataset[T, O, D, S] now becomes Dataset[T]. Similarly, Iterator[T, O, D, S] now becomes DatasetIterator[T].
• First steps towards type safe auto-differentiation.
• Removes the dependency on spire.

NOTE: This does not compile yet, but I'm working actively on getting it ready.

[eaplatanios]

@sbrunk You may be interested in this. It's one of the big changes I mentioned some time ago.

[sbrunk]

Agreed. This is a great simplification. :)