trigger persistenced in cos-extensions and not in the gpu-installer

This reverts the change where persistenced was moved to be triggered from the cos-gpu-installer.
This causes issues on GB300. Probably because of the very slow boot on bare-metal.

This CL reverts the logic for persistenced done in these CLs:
1. https://cos-review.git.corp.google.com/c/cos/tools/+/149184
2. https://cos-review.git.corp.google.com/c/cos/cos-extensions/+/149046
3. https://cos-review.git.corp.google.com/c/cos/tools/+/154503

BUG=b/542405340
TEST=Build and run on a4xmax pod. Check cuda and that:
sudo /home/kubernetes/bin/nvidia/bin/nvidia-smi -q | grep -E "Persistence Mode|Attached GPUs|Product Name"
shows Persistence Mode: Disabled
Then running:
sudo /home/kubernetes/bin/nvidia/bin/nvidia-smi -pm 1
which turns Persistence Mode: Enabled
RELEASE_NOTE=NONE

cos-patch:bug
Change-Id: Ib185b94b0cf02c88bc1b1788d634d59da5ff4f46
Reviewed-on: https://cos-review.googlesource.com/c/cos/tools/+/179466
Cloud-Build: 228075978874@cloudbuild.gserviceaccount.com <228075978874@cloudbuild.gserviceaccount.com>
Tested-by: Miri Amarilio <mirilio@google.com>
Reviewed-by: Derek Taylor <ddtaylor@google.com>