Fix N150 llama3 demo CI tests to properly save perf information to superset

[mtairum]

Final change to these CI tests after another back and forth:

Before we were running Llama3 on both N150 and N300 as intended. However the N150 functional pipeline does not upload the perf data to superset, nor does it run on the perf VMs.

This PR changes that.

Instead of running Llama on N150 pipeline, we run them twice on the N300 perf:

• First time we run them with FAKE_DEVICE=N150 which forces a single-chip execution and saves the N150 perf to superset.
• Second time it runs normally, on a 2-chip N300 and saves the perf to superset.

To make sure this is the last change we are running the CI test and will make sure all the data is uploaded before committing this.

• https://github.com/tenstorrent/tt-metal/actions/runs/11918493617