FEAT: Add round-robin adversarial target endpoints

[Adrian Gavrila (adrian-gavrila)]

Summary

• add environment configuration for up to three adversarial chat endpoints
• publish multiple configured endpoints as a RoundRobinTarget under the existing adversarial_chat name
• preserve single-endpoint behavior and direct access to individual members
• keep explicitly configured adversarial endpoints out of inferred auto-grouping
• cover single, multiple, noncontiguous, incompatible, and auto_group=False configurations

Why

This allows adversarial workloads to rotate across regional deployments without requiring scenario changes.

Backward compatibility

Existing users do not need to update their configuration. When only the original ADVERSARIAL_CHAT_* endpoint is configured, adversarial_chat remains a direct OpenAIChatTarget; a round-robin target is created only when multiple endpoints are available.

Users configuring regional deployments with different deployment names should assign them the same ADVERSARIAL_CHAT_UNDERLYING_MODEL* value so PyRIT can verify that they are behaviorally compatible.

Testing

• new and existing unit tests in affected code paths pass
• pre-commit hooks pass