Enhancements to batching

[neworderofjamie]

The batching system does a few things which make sense in typical ML use cases but, reduce it's usefulness as an efficient parameter-sweeping tool:

• VarAccessDuplication::DUPLICATE variables are currently initialised once and the same value is copied across all batches. We probably should 'allocate' num things * num batches RNG streams and decide which ones to actually use on a variable-by-variable basis (controlled by flag in Models::VarInit perhaps)
• Currently connectivity i.e. ind, rowLength etc have the equivalent of VarAccessDuplication::SHARED as this fits with typical ML uses of batching. However, as long as max row length etc stays the same across all batches, no reason not to make this more flexible