CommandQueueMT: Fix flush re-entrancy

[RandomShaper]

CommandQueueMT re-entrancy was broken two-fold:

• First, the collaboration with WorkerThreadPool (thread_enter/exit_command_queue_mt_flush()) doesn't have a concept of stack or ref counting.
• Second, the read offset was incremented too late, so a re-rentrant call would re-read the flush command that has already been processed.

This PR fixes re-entrancy issues by just early-returning. The result will be the same as if re-entrancy had been detailedly fixed so it was possible without issues, but wihout getting a deeper stack.

Fixes #90363.

[clayjohn]

Lets go forward with this for now. It does fix the crash and most likely will work fine in most cases.

But we need to decide if the implied ordering of commands resulting from this PR is what we want (I think it is).

[akien-mga]

Thanks!