Nd4j: Provide faster implementation of `diagonalEmbedding` in `Nd4jTensorAlgebra`

[CommanderTvis]

Postponed because of #423.

[grinisrit]

It is not a question of performance only, it has to be also compliant with https://deeplearning4j.org/api/latest/org/nd4j/autodiff/functions/DifferentialFunction.html for future use.

[grinisrit]

To keep it there for ref: https://deeplearning4j.org/api/latest/org/nd4j/linalg/api/ops/impl/shape/Diag.html