memory batching

[Shayan-P]

Right now, we are loading the whole Stabilizer to the GPU memory. Depending on the user's GPU, it might not always fit.
A 2000 gate Pauli_Frame with 2^19 trajectories takes about 2GB of memory.

so we might want to split the memory into multiple batches and load as we need.

Or

we can pin the memory (I think this will allow GPU to use it without having to load it).

[Krastanov]

I suspect the wins would not be very big compared to just using the GPU for small batches like this. It would be interesting to have this, but no need to put it high on the priority list.