Explore removing std::vector mData; completely from Tensor in favour of always storing data in hostVisible buffer memory (TBC)

[axsaucedo]

As followup of #136, opening issue to focus on researching phase 4 specifically, which was presented as following:

4) Remove std::vector mData; completely from Tensor in favour of always storing data in hostVisible buffer memory (TBC)

After doing some resarch I was able to confirm:

• All GPU devices will always have both HostVisible and DeviceVisible memory available for use
• Performance of GPU hostVisible memory access from CPU is relatively efficeint (even if not as efficient as RAM), this AMD document mentions "up to 64 GB/s peak theoretical transport data bandwidth from CPU to GPU per card." - still needs to be further confirmed
• It is possible to leverage std-compatible containers such as std:span-like containers (or msft's gsl/span) to have a container with non-owning reference to the host-visible memory

Due to these points, it may be possible to completely remove the std::vector mData attribut completely whilst still keeping the best of both worlds - namely 1) access to the memory that was added to the GPU, and 2) optimum memory usage (mainly by not storing an extra copy of the data).

This point would still need to be explored further as tehre are some nuances in regards to how this would be implemented in a memory-safe way, as well as ensuring that the python interface can be exposed efficiently, as ideally it can be built in a way in which it can be passed to and from python with minimal amount of copies (as it's not clear what the support for non-owning containers like span is)