With the switch to using the primary CUDA context it became possible
for peer access between CUDA devices to already have been enabled for
that context, either by a previous Cycles session or third-party library,
thus causing the call to `cuCtxEnablePeerAccess` to return
`CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED`. This is not a failure
state however, so just needs to be handled like a success return value.
Pull Request: https://projects.blender.org/blender/blender/pulls/120255
33d7fa8cb3