Adds a mechanism to manually promote constants to inputs in Lazy Tensor.

[bgogul]

Adds a _concreteInputLazyTensor property to the tensor-related types in the library. This provides temporary mechanism to promote constants to inputs manually in trace. Here is an example use case:

import TensorFlow

let a = Tensor<Float>(10.0)
let b = Tensor<Float>(5.0)
let w = a._concreteInputLazyTensor * b
print ("\(w)")

This will result in the following trace for w:

lazyTrace_3(%0: float) -> (%2) {
  %1 = Const[dtype: float, value: 5.0]()
  %2 = Mul[T: float](%0, %1)
} 

Note that the slot for a got promoted to an input. This is added to experiment with the lazy tensor implementation and is mostly intended to be used within the TensorFlow module. If this turns out to be useful, we should provide a cleaner user-visible API.

This PR also adds a helper file for tests related to lazy tensor.

[pschuh]

Looks fine. Could you include a motivating example in the description?

[saeta]

+1 for documenting a motivating example. Also, could we start with this being _-prefixed? I'm making a few guesses, but we might want to express these concepts in a slightly different way. I'd like us to carefully think through the API design space here. Does that make sense? Thanks!

[rxwei]

If this protocol is internal, it need not be underscored. The underscoring rule only applies to implementation-detail types that had to be made public.

[rxwei]

The protocol is internal. Why does this requirement need to be public? Same for other occurrences of public.

[bgogul]

@pschuh @saeta @rxwei I scaled back this PR to publicly expose only what I needed. I also underscored it.

I moved the _LazyTensorCompatible protocol to a test helper.I don't think it is useful outside of the tests for the time being. If we have a need, we can reassess and design the API properly.

[pschuh]

Constants in the example don't match up.

[bgogul]

Constants in the example don't match up.

Thanks for flagging it. Fixed now.

[rxwei]

Hmm is there a reason this needs to be public?

[bgogul]

I expect that it will only be used within the library. However, I wanted it to be available for experimentation in the user models if needed.

[rxwei]

How and why should a user be expected to use this?

[rxwei]

Just wanted to be careful exposing public APIs. I think we need to make sure public APIs have at least some justification of use cases before they are exposed.

[bgogul]

The use case is the one in description where we want to force a constant to be an input to an extracted trace. Ultimately, we may not need this API. However, in the meantime, I want to have this knob while we develop and improve the lazy tensor implementation.

[rxwei]

Oh I see. Sorry for missing parts of the PR description! And thanks for explaining :)