Provide a simulation mode that guarantees no out of memory errors

[PabloAndresCQ]

This will require keeping track of memory currently allocated in GPU, and ask cuTensorNet how much memory the next primitive (contract or tensor decompose) will require.

[PabloAndresCQ]

Proper support for this will be possible after the next release of CuTensorNet (see NVIDIA/cuQuantum#156). A hack is possible at the moment, but perhaps it is not high priority.