FEAT GCG algorithm and AML pipeline

[blakebullwinkel]

Description

This PR adds the GCG algorithm, which can be used to generate adversarial suffixes that jailbreak language models. The GCG code is mostly the same as that in the original repo, with a few changes including:

• Added support for Phi-3-mini
• Logging using mlflow (functions added to pyrit/auxiliary_attacks/gcg/experiments/log.py)

This code been added to a new high-level directory pyrit/auxiliary_attacks, which will contain the code for attacks that do not fit into the core PyRIT functionality. This PR also adds a doc/code/auxiliary_attacks directory with two demo notebooks:

• 1_auxiliary_attacks.ipynb: explains what we mean by "auxiliary attacks" and shows how to send a GCG suffix to a Phi-3-mini AzureMLChatTarget
• 2_gcg.ipynb: provides the full AML pipeline to generate GCG suffixes including environment creation and job submission

Several additional dependencies are required to run the GCG algorithm. These have been added as optional dependencies under gcg in the pyproject.toml. We also update the github build pipeline to install these dependencies along with the dev dependencies so that we can continue to support this functionality in the future.

Tests and Documentation

Given that the GCG code runs independently from the rest of PyRIT, we decided not to add tests for it. Documentation has been added to doc/code/auxiliary_attacks.

[rdheekonda]

Great work, @blakebullwinkel !