Raw versions of tensors

[mortendahl]

The current implementations of storage modules is using eg HostFloat64Tensor::from to lift a (raw) ndarray tensor into a Moose tensor (and in turn a Value), yet these From impl do not have placement info and hence the loaded Moose tensors have fake placement information.

This issue is to discuss whether we want

• the storage module to be aware about the placement by passing it in through the load/save kernels
• or instead only ask storage modules to produce some notion of RawValue that do not include placement information, and then add logic to the load/save kernels to convert from eg RawValue to Value by adding placement information
• we should only send raw values, not placed values