[WIP] Start doc on data attributes / data locators

[mwiebe]

No description provided.

[mrocklin]

Is this roughly how DyND thinks?

Sometimes we use datashape with storage mechanisms like JSON or SQL or compressed HDF5 for which this kind of thinking probably doesn't make sense. How do we handle this mismatch?

[mwiebe]

This is along the lines of how DyND thinks, yes. In DyND, each type is associated with an array metadata layout, and gets a little bit of memory for that in each array object.

An array of type strided * pointer[int32] has a bit of array metadata information for strided with size and stride data items, immediately followed by some for pointer with an offset and a reference to track ownership of the block of memory all the pointers are pointing into, and then nothing for the int32.

By making the dattr be json, it's going to be less performant but more portable. Things like the reference to track memory ownership for the pointer type probably don't belong in here either.