Fix alignment and locking issues with CommandQueueMT

The allocations of commands in CommandQueueMT weren't aligned. This commit aligns all accesses on 64bit boundaries regardless of target platform. This ensures that all types are aligned.

Lock-wise the semaphores were maked as usable when the command had ran but not when the synchronous stub had finished with it. This lead to a race condition where sometimes the semaphore got reused before it was waited on. We now mark the semaphore as free only once we're done waiting on it.

This fixes #21725 (closed)