Add Modal.com GPU server for nnInteractive v2 + fix macOS pip crash

[pieper]

Summary

• server/modal_app.py — new Modal.com deployment that serves nnInteractive v2 as a drop-in replacement for the local server at port 1527. Point the Slicer extension at the Modal URL instead of http://localhost:1527.
• SlicerNNInteractive.py — fix macOS crash when installing dependencies: slicer.util.pip_install creates Qt dialogs from a background thread, which crashes on macOS because NSWindow must be created on the main thread. Replaced with subprocess.check_call running pip directly on the main thread.

Modal server changes from the upstream v1 server

• nninteractive upgraded 1.0.1 → 2.4.2
• use_pinned_memory removed (replaced by interactions_storage='auto' in v2)
• interaction_bbox added to lasso/scribble calls (~1000x speedup by limiting inference to the drawn region)
• /health and /reset endpoints added
• Model weights cached in a persistent Modal Volume (no re-download on container restart)
• container_idle_timeout=360 keeps the container alive during a session
• allow_concurrent_inputs=1 serializes requests to protect in-memory session state

Backward compatibility

The HTTP API shape is identical between the v1 and v2 servers — all existing endpoints (/upload_image, /upload_segment, /add_point_interaction, /add_bbox_interaction, /add_lasso_interaction, /add_scribble_interaction) use the same request/response formats. The Slicer plugin works unchanged against either server.

Test plan

• modal serve server/modal_app.py starts without errors
• Point Slicer extension at the ephemeral Modal URL; upload image and confirm all prompt types (point, bbox, lasso, scribble) return correct segmentations
• modal deploy server/modal_app.py produces a persistent URL
• Confirm macOS no longer crashes during first-run dependency install

🤖 Generated with Claude Code

[pieper]

Addresses #96