feat: FGSM robustness tool, Grad-CAM interpretability tool, and async multi-tool orchestrator

[sisodiyaanand]

Summary

Adds the first deterministic tools (per AGENT_TRACKER.md, which currently lists no implemented agents) for adversarial robustness and interpretability evaluation, plus a generic async orchestrator to run heterogeneous tools concurrently.

• FGSMAttackTool — FGSM adversarial perturbation, measures prediction flips (extends AbstractBaseTool)
• GradCAMTool — Grad-CAM saliency heatmaps (extends AbstractBaseTool)
• AsyncAgentOrchestrator — runs multiple tools concurrently via asyncio.to_thread, with honest per-tool error reporting instead of silent failure (extends AbstractBaseAgent)

All three have dedicated tests; full suite passes (5/5, including the existing test_smoke.py).

Open question for maintainers

These tools currently take raw torch.Tensor I/O rather than going through the LLM-driven DLensBaseAgent. I used the deterministic AbstractBaseAgent/AbstractBaseTool pattern instead, since tensor I/O doesn't map cleanly to JSON-serializable LLM tool calls. Wanted to check if this is the right pattern, or if there's a preferred way to eventually bridge this into the conversational/LLM agent layer.

I'm an MSCS student at Georgia Tech, interested in this as a potential GSoC 2027 contribution area. Happy to iterate based on feedback — opening this as a draft to start the conversation rather than requesting immediate merge.