Add support for Groq and Parallel Supervisors

[mrfire15]

Inspired by the paper, I experimented with two strategies:

• Early Stop: Run multiple supervisors in parallel and terminate the remaining ones as soon as the first completes.
• Aggregator: Run supervisors in parallel and aggregate their results afterward to potentially reduce the total number of iterations.

Additionally, I integrated support for Groq to test these strategies for free (though rate limits are quickly encountered :( ).

Observations:

• When multiple supervisors execute in parallel, they tend to fetch highly similar data, leading to redundancy and wasted computation.
• In the aggregation approach, the merged results often contain overlapping information, offering limited added value.

Key difference from the paper:
In the referenced paper, the multi-agent system included a planner component that generated and coordinated a complete plan(can be diverse) for the agents to follow. In contrast, open_deep_research doesn't need much planning. It consists mainly of a supervisor that spawns several research agents in parallel (via prompts), collects their outputs, and triggers another iteration if more data is needed.

Without a diverse and plan-driven multi-agent setup, both the early stopping and aggregation methods don't seem to give much benefit in this context.