Add convenient write interface to SharedTensor

[MichaelHirn]

For writing to the memory copy of a SharedTensor, we rely on the write_to_memory helper function, explicit synching and other boilerplate. This is bloating the tests and makes adoption quite hard and not fun at all. This got also pointed out at the NN Plugin in this PR.

A simply interface for writing (initial) data to a SharedTensor is missing. I would propose to keep the implementation close to e.g. numpy or torch with well known helpers such as zeros, ones or random, which are quite convenient for a lot of tasks - e.g. ANN/ML related stuff.