feat(inferno): extract inference orchestration into core/inferno/orchestrator.py

[slomin]

Summary

• Extract health probing, readiness state machine, process restart coordination, inference tick decision logic, mmproj/no-mmap resolution, and activation runtime prep from core/main.py into core/inferno/orchestrator.py
• Orchestrator accepts primitives, dicts, and callbacks — no FastAPI dependency. Launch arg assembly and runtime env building stay in main.py where RuntimeConfig is available. The tick receives launch callbacks so main.py controls process spawning while the orchestrator owns the decision logic.
• Rollback hardening: failure counter tracks runtime family alongside model path, tick skips spawning when runtime switch lock is held, activation delegates compatibility pre-check to prepare_activation_runtime before any state mutation
• 43 new orchestrator unit tests including pinpoint regression tests for termination failure propagation and missing-script readiness downgrade

Test evidence

$ uv run python -m pytest tests/unit tests/api -q -n auto
895 passed in 10.28s

$ npx playwright test tests/ui/ --reporter=dot --timeout=15000 --workers=75%
173 passed (38.8s)

Pi QA

Required — model/runtime switching must be validated on real hardware.

Closes #297