Approach to dependencies

[KOLANICH]

Currently this package lacks support of nodes. But numpy-ml has them. While it in some part duplicates scikit-learn, CNNs are explicitly out of scope all numpy, scipy and scikit-learn and are in scope of numpy-ml.

Also it'd be nice to make use of installed ML frameworks and other libs, like OpenCV, if it brings benefits.

So the proposal

1. Lift a restriction of the runtime depending on numpy only
2. Implement the missing operators using numpy-ml and state in the docs it may be needed fir some ops.
3. Detect presence of libs and use them for some operations if they are present. It'd require some benchmarking to determine which lib it is more optimal to use in each case.