Implement BatchNormalization functions from cuDNN in juice

[colepoirier]

What does this PR accomplish?

• 🦚 Feature

Closes #145.

Changes proposed by this PR:

• [ ] Adding the necessary wrapper functions of the ffi::* cuDNN BatchNormaliztion functions to rcudnn/cudnn/src/api/.
• [ ] Adding the corresponding functions to coaster-nn.
• [ ] Adding the corresponding layers to juice.

Notes to reviewer:

I think that at least one additional new function https://docs.nvidia.com/deeplearning/cudnn/api/index.html#cudnnDeriveBNTensorDescriptor needs to be added in order to add the BatchNormalization functions to juice as: "This function derives a secondary tensor descriptor for the batch normalization scale, invVariance, bnBias, and bnScale subtensors from the layer's x data descriptor."

I am not sure how to proceed with this in terms of how to properly add this/implement it in rcudnn.

📜 Checklist

• [ ] Test coverage is excellent
• [ ] All unit tests pass
• [ ] The juice-examples run just fine
• [ ] Documentation is thorough, extensive and explicit