Add DS-Chat CI workflow

[lekurile]

This PR adds a DS-Chat CI workflow for testing Step 3 in the DS-Chat pipeline.

This test sweeps across the following parameters:

Zero Stage:    2, 3
Hybrid Engine: True, False
Offload:       True, False
LoRA:          True, False

The models used for this test are:

Actor:  facebook/opt-125m
Critic: facebook/opt-125m (trained in DS-Chat Step 2)

Each configuration (16 total) runs through a limited number of Step 3 non-overflow training steps (i.e. steps where neither actor nor critic overflow) and saves the actor/critic models. Assertions are used to check if the training pipeline executed correctly and if the actor and critic models were saved properly.

The test is scheduled to run nightly at midnight.