Changes:
- Accepts primitive types (e.g. types with no fields).
  This allows broadcast to produce SoA arrays starting
  from something like `SoA{Float32, 2}`.
- Supports multiple storage array types. Most notably
  `CuArrays` and `CuDeviceArray`.

In this process the broadcasting was updated to work and most of the
tests were updated to run.

Here we have updated the README to include the new way of constructing
`StructOfArrays` and have removed dated information.

This adds basic support for broadcasting a `StructOfArrays` with `CuArray`
storage.  I could not figure out how to reuse the functionality from
GPUArrays without copying it here.