Clean up unused ToDevice sync and associated memory leak (#71744)
- storing a list of results[] inside timed function was a workaround to
enable keeping tensors alive long enough to use them for to_device
sync
- we don't use this sync method anymore, and the infra for it is costing
more than its worth.
- the small memory leak this fixes should be isolated to eval models,
and should be predictable increments of the size of the tensor
returned by the eval function, which were still cleaned up at the
conclusion of an amortized experiment, just kept alive during the
internal loop. This isn't expected to be a big deal for OOM issues,
but made noise that complicated debugging other OOM issues.
