Added support for 'Tensor.batchGathering(atIndices:)'.

[eaplatanios]

@rxwei this also makes use of Tensor.withoutDerivative where Scalar is not differentiable. How do you think we should avoid these uses here?

I also have a more general version of this that requires control flow AD so instead of committing it as non-differentiable but more general function now, I'll leave that one for later.

[eaplatanios]

@rxwei I removed Tensor.nonZeroIndices(), as discussed and added a noDerivative global function helper. I could not name it withoutDerivative because that causes compilation errors due to a conflicts with the existing withoutDerivative. This passes all tests now.

I'd prefer we call it withoutDerivative and remove the existing withoutDerivative method though from Differentiable.

[eaplatanios]

Nvm the Tensor.nonZeroIndices() edit. It's actually used by Tensor.gathering(where:) and has already been merged into master.

[eaplatanios]

@rxwei I updated this PR so that all tests pass, but there is one issue that came up. I had to use Swift.withoutDerivative(at:in:) because if I don't use the module to disambiguate the compiler tries to use the instance method of Differentiable and fails. Not sure why though since the signatures are different. This seems like a compiler bug.

[rxwei]

Swift.withoutDerivative(at:in:) sounds good. This is a bug which I was trying to fix with swiftlang/swift#25316, but it needs a more work to get there.

[eaplatanios]

Right, I also noticed the same problem with the new elementary function support. Thanks for the update! This should be ready to merge then, along with the few other PRs I updated today. I also ran CI tests on them.

[rxwei]

Just got back from a conference. Will get to these PRs tonight!
For this particular one, feel free to merge when the comment above is addressed.

[eaplatanios]

Thanks Richard!