Motivation: deeplearning4j/deeplearning4j#8582

The main challenges here are of course the numerical ranges vs. java types, in terms of API for getting/setting values.
uint32 isn't so bad, as it fits in a long (like UByteIndexer uses int)
However uint64 can clearly represent values larger than long.

Not sure what the best solution for uint64 is, but some options are:

• use long (ignoring the overflow)
• add a new/custom Number class (uint64 stored as a long, converted as required)
• BigInt (not exactly a low overhead solution however)