Adds a declarative MetaInstance system for managing model instances with
automatic placement, retry logic (max 3 attempts), and lifecycle
management via a reconciliation loop.
- Process managers for instance health, meta-instance lifecycle, and
node timeout detection
- Reconciliation engine driving state transitions and cascading deletes
- Dashboard UI for creating/managing MetaInstances with node selection,
sharding config, retry status, and error feedback
- JACCL SideChannel integration for distributed inference coordination
- Comprehensive test suite (25+ edge cases)
Split from original #1519. Independent bug fixes extracted to:
#1547 (misc fixes), #1546 (JACCL sidechannel), #1582 (download
detection), #1580 (RDMA warning)
Co-Authored-By: Claude Opus 4.6 <[email protected]>