radv: Eliminate zero commandbuffer submissions.

A zero commandbuffer submission can be expensive with the global BO lists.

There is some room for optimization if we can use a timeline semaphore for the implementation (as I don't think amdgpu can see through the sync file tricks to submit sumissions to the HW early) but it shouldn't be worse than the zero cmdbuffer submit (since the lookahead was 1 anyway and the empty submissions took ~0 time on the GPU). So for now I've kept it simple.