Switch to multiGPU AWS instance for regression testing

[elezar]

Bugs such as those fixed by #1355 require multiple GPUs to trigger. We should switch our testing instance to one with at least two GPUs.

We currently use g4dn.xlarge instances. Switching to a g4dn.12xlarge or g5g.16xlarge
will give us multiple GPUs (although the latter may require additional changes since it's a arm-based system).

Ideally we would have mutli-GPU specific tests that can be filtered so that we don't run these on every PR, but do run them on main or the latest release-* branches.