chore: use kepler_vm metrics as energy source

[rootfs]

When training models with kepler_vm_package_joules_total or similar, we use the trained_power_model as energy source, this fixes the issue of getting the proper energy from prom metrics.

Tested with the following command:

hatch run python cmd/main.py train --vm-train -e trained_power_model

[KaiyiLiu1234]

lgtm. Let's give this a try on metal ci.