Provide further granularity on handling staging tensors

[axsaucedo]

Currently staging tensors are created by OpCreateTensor but they are kept for the life of the OpCreateTensor. Should explore ways to expose more granular functionality, whereby the mStaging tensors may only be needed once, for example, in the case of input tensors. There are other cases where the staging tensors would be expected to be used many times, in case that ther input tensor is to be updated multiple times from the host (to avoid creating a staging tensor every time). This may require potentially moving the staging tensor ownership into the Tensor class itslef. This would mean that the ownership for the vulkan memory of this tensor would still be owned by the life of the OpCreateTensor, but it would still be possible to have other functions like an OpMapData (as per #39) which could still leverage the staging tensor (whcih can either be already existing or created/destroyed in the single operation.