feat(ccu): add direct runtime engine

[Kur0x]

Summary

• add the direct CCU runtime lifecycle and public-runtime launch boundary
• add the direct orchestrator on top of the install-provider stack
• retain TileXR-owned types and avoid hcomm or HCCL private APIs

Stack

4 of 9. Depends on codex/ccu-stack-03-install.

Validation

• git diff --check
• focused direct-orchestrator and RA-loader unit tests
• 36 tests passed; 13 CANN or later-layer cases skipped

Scope

This PR adds the engine only. Communicator and collectives integration follows in the next PR.

[Chand1erLiu]

Please make this owning runtime RAII by adding a destructor that calls Shutdown() (and keep it non-copyable). Once initialized it owns registered memory/token handles plus channel/CQ/QP/import handles; allowing the object to leave scope without an explicit shutdown leaks those resources, and TileXRCcuHccpLoader cleanup does not release them.