v3dv: fix double free error when releasing sems_info resources

Why do we have a double free exactly? When we have a wait events job, we supposedly make a new copy of the semaphores for that, that we then destroy when the thread ends, so I presume the double-free means that when the we are done waiting for the events and before the thread wait ends and we call:

VkResult result = queue_submit_job(queue, pjob, info->sems_info, NULL);

for the remaining jobs, some of those jobs are holding a reference info->sems_info and eventually free it, which I don't think is supposed to happen. So what job is trying to do that exactly? I think we need to narrow it down to understand what is happening exactly.

I was thinking the case where we have 2 wait for event jobs in the same command buffer, but I think that should work fine: the first one spawns the wait thread and copies the semaphores, then when the wait is done, we execute the second wait for events job in the same wait thread (it won't copy the semaphores) and the semaphores are freed once after the threads ends, so there should not be a double-free.

From what I tracked in this test case (using the commit that introduced the double free but before multisync support):

• We have two wait_event_cpu_job in different command_buffer_batch (therefore different command buffers)
• In the first wait_event_cpu_job from the first cmd_buf_batch, (cpu.event_wait) -> info->sems_info == NULL and we copy semaphores and spawn a wait_thread, as expected.
• In the second wait_event_cpu_job from the second cmd_buf_batch, info->sems_info is not NULL, we copy semaphores again, spawn another wait_thread (because we are in a different cmd buffer) and lose the sems_info reference.
• Finally, the wait_thread to both events finish, the first free the sems_info and the second crashes by trying to double free it.

I'll dig a bit more.