Support for "combining" operators

[blever]

The ability to combine multiple arrays into a single one can currently only be achieved using zip/zipWith. Of course, these arrays can only combine 2 arrays at a time, so in general it would be useful to have operators for combining more than 2 arrays in more elaborate, but structured, ways.

Some requirements could be:

• 'combine' operator:
  • a generalised array combining operator
  • sum of input arrays sizes is equal to output array size - i.e. no elements are lost or duplicated
  • no permutation is performed on elements within in input array - i.e. output array elements are contiguous with respect to their source input array
  • input array elements must all be of the same type and shape (shape would need to be a run-time check)
  • the combination does not need to preserve dimensionality - e.g. multiple 1D arrays could be combined to produce another 1D array (concatenation), or a 2D array (stacked) or even "maybe" a 3D array (stack-stacking?)
• 'append' operator:
  • a specialised array combining operator
  • two input arrays - place one array at the "end" of the other
  • input array elements of the same type
  • the higher dimensions of the input arrays must have the same extent

In addition to fulfilling the need of a common pattern, combining operators would allow for further backend optimisations:

• input arrays to a combine operator would not require intermediate writing to memory - they can be written directly (by their producer) to their location within the combined output array
• on architectures such as Fermi, the generation of the input arrays can be done in parallel by using separate streams - 'combine' in this case acts as synchronisation barrier until all computations are complete