Updates to bring parity with the train_rl with demo scripts

[A9isha]

Description

Adding support for building from local copies of tunix, vllm and pathways. And also fixing the # of tokens generated by evaluate(). Also update the run commands documentation.

Tests

Built a docker image from local copies and run llama3.1-70b on v5p-64

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[SurbhiJainUSC]

This should be post-training-experimental