storage: supervise startup and config application

[jveski]

• gate shard serving on phase-B activation and initial disk publication
• precompute and share loaded config projections and route snapshots
• preserve live resources on failed reloads and surface disk and peer reconciliation failures
• supervise shard termination and fail coherently on indeterminate config application

[jveski]

AI review: This comment was generated by an AI agent, not by the PR author.

• Blocker: cmd/unbounded-storage/src/main.rs:922-972 loses failed forced frontend rebuilds. When a cache binding or backend stripe size changes, forced_frontend_updates rebuilds an otherwise unchanged frontend. If that replacement fails, last_frontends retains the old live driver, but lines 960-964 treat the unchanged frontend spec as proof that the desired binding applied and overwrite last_bindings with it; a successfully replaced backend similarly advances last_backends. On the next apply of the same desired snapshot, frontend_rebuild_ids therefore sees neither binding nor geometry drift, skips the frontend replacement, and reports success while the live driver still uses its old backend/cache/stripe geometry. Preserve the old auxiliary state for failed/deferred forced updates and add a retry regression covering a frontend build failure after binding or stripe-size drift.