deep desync analysis updated

Changed

• Breaking: EntityGrainOptions.DeepDesyncEnabled (bool?) → DeepDesyncMode { Off, PerPlayer, Forced }. Migration: null → PerPlayer, true → Forced, false → Off.
• Off now outranks a client request. It used to be OR-ed with the per-session toggle, so a client could switch CRC computation back on against the documented kill switch.
• Under PerPlayer the flag lives on DesyncReportGrain — set by SetDeepDesyncAsync or admin tooling, applied to the running session immediately, and kept across reconnects.
• The client builds a patch tree only while the analysis is on. A build carrying [MetaServiceImpl(DeepDesync = true)] used to allocate and hash one on every call whether or not anyone was looking.
• RpcCall.DeepDesyncRequested → DeepDesyncActive. Wire ids unchanged.

Added

• Broadcasts carry a patch CRC whenever the server already built a patch tree, and an analysed client verifies its replay against it — a divergence caused by another player's call is now visible. ServerPatch / ServerReplace broadcasts stay unchecked: the client computes nothing there.
• SessionConnectResponse.DeepDesyncActive / INetwork.DeepDesyncActive — the client is told the verdict at connect.
• Startup logging: the silo lists the services able to report, the client logs the session verdict plus which of its own services are covered.
• SetDeepDesyncAsync returns false when the silo cannot honour the request (Off, or Forced asked to switch off) instead of accepting and ignoring it.

Fixed

• [MetaServiceImpl(DeepDesync = true)] on a service declaring [ServiceConfig] failed to compile — the _PatchTracked copy emitted only the legacy Config accessor. Affected any force-patch-able service on the 0.33.0 config API, not just deep desync.
• DeepDesyncEnabled's docs claimed null fell back to the per-service attribute and false was a kill switch. Neither held: null meant off, and false was overridable by the client.