Add 16-bit float tensor types

[karllessard]

This PR shows how to make use of DataLayouts to implement new tensor data types not directly supported by Java primitive types. It adds 2 new tensor types to the Java bindings: TFloat16 and TBFloat16. The former implements the IEEE-754 half-precision floating-point specification while the latter is for truncated 32-bit floats, called "Brain float-16".

In addition, you'll notice that the PR also include a lot of javadoc updates, which is a work continuously in progress.

[Craigacp]

Javadoc looks good.

Is it worth adding a method to the BFloat16 and Float16 interfaces that accepts a Tensor<Float16> and returns a Tensor<Float32>? That'll make it simpler for users to convert the result of a computation into something they can easily use from Java.

[saudet]

Intel capitalizes it "Bfloat16" so it might be better to do it like that here as well?
https://software.intel.com/en-us/download/bfloat16-hardware-numerics-definition
There are also more instances of "Bfloat16" than "BFloat16" in TensorFlow's main repository,
but it's not too clear cut...

BTW, again, why do we need to prefix all those types with "T"? What's wrong with just Float16, etc?

[Craigacp]

I think the prefixing helps to distinguish them from things that implement Number. At the moment they are more like tag types to make things flow through the type system properly rather than data carriers.

I've no opinion either way on BFloat16 vs Bfloat16, whatever is more prevalent in the community is fine by me.

[karllessard]

For the prefix, please see my reply in the group thread as this question has been asked in many different places :) https://groups.google.com/a/tensorflow.org/d/msg/jvm/Zp362xTQahc/Ho6iUEJfAgAJ

For the BFloat16 vs Bfloat16, I'm fine with both as well so let's follow @samuel suggestion then, I'll rename it.

[karllessard]

What about TUInt8 vs TUint8? Again, both are used in TF core.

[karllessard]

Haha, worst, all of QUInt8, QUint8 and Quint8 are present in TF.

I guess if we go with TBfloat16, we should apply the same pattern of capitalizing only the first letter for all types. It looks like that is what the Go client is doing. So if everyone is ok, I'll make the changes for TUint8 as well.