[8.3.2] ALCF Agent

[davramov]

This PR introduces an agentic workflow using Academy Agents that runs reconstruction and segmentation on ALCF.

• Utilizes the Globus Compute Multi-user Endpoint on ALCF. This is great because it is maintained by ALCF, and does not require us to set up our own compute endpoint. You can pass in the queue/node/etc requirement as a parameter for the job.
• Utilizes the alcf-ai inference service:
  • A vision LLM call is made that assess the quality of a reconstructed slice to determine if it is good enough to send for segmentation
  • If that check completes, the reconstruction agent hands off to the segmentation agent, which uses the ALCF SAM3 inference service. This is also great because the service is "warmed up", so time is not spent waiting for the model to load.

Video demo:
https://drive.google.com/file/d/15Rv82W5HZhGXT1GRqAEHLCx-C2n-m7Qi/view?usp=drive_link

Notes:

• still needs to be rebased
• uses Yadu's skeleton for setting up agents: https://github.com/als-computing/splash_flows/blob/agentic_rework/orchestration/agentic/skeleton.py
• potential to add additional features...
  • What if the agent runs reconstruction of 1 slice (i.e., contrast/sharpness), assesses with a vision-enabled LLM, and iterates on the parameters (e.g., turn on/off phase retrieval) to find a good setting before running full recon
  • What if the agents read the metadata from the h5 file, such as the abstract, and an LLM helps pick a good prompt for segmentation
  • For fine-tuning workflows, we could think about the agents capturing the center-slice from reconstructions, and sending it to Tiled/segmentation app