End-to-end tests for torch_xla llama and mixtral

[tengyifei]

Fixes #52

E2E test design

The workflows in e2e_test.yml does a few things:

• Set up gcloud credentials from a Service Account key managed in repo secrets.
• Install torchprime.
• Test tp use and point it to an XPK cluster hosted internally.
• Test tp run on a few models.

After kicking off the training of some models, it starts a parallel job for each
model, and runs a few checks. This is implemented in
reusable_e2e_check.yml:

• Stream the logs.
• Check workload exit code.
• Check for specific log strings that indicate training success.
• Check that there is a profile .pb file.