v3d: fix memory barrier implementation

@anholt: I noticed that a job which included SSBO writes was not being flushed after the application called glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT). The job was still being flushed when the application mapped the buffer though, but going by the comment in that function I think the intention here was to flush at the memory barrier, right? Also, the comment in that function says that we want to flush jobs writing to ssbos or images, but then it didn't seem to be doing that either.