r600g: use WAIT_3D_IDLE before using CP DMA

I broke this with 7948ed1250 for r700 at least.

src/gallium/drivers/r600/evergreen_hw_context.c

@@ -123,6 +123,7 @@ void evergreen_cp_dma_clear_buffer(struct r600_context *rctx,
 
 	/* Flush the cache where the resource is bound. */
 	r600_flag_resource_cache_flush(rctx, dst);
+        rctx->flags |= R600_CONTEXT_WAIT_3D_IDLE;
 
 	while (size) {
 		unsigned sync = 0;

src/gallium/drivers/r600/r600_hw_context.c

@@ -636,6 +636,7 @@ void r600_cp_dma_copy_buffer(struct r600_context *rctx,
 	/* Flush the caches where the resources are bound. */
 	r600_flag_resource_cache_flush(rctx, src);
 	r600_flag_resource_cache_flush(rctx, dst);
+        rctx->flags |= R600_CONTEXT_WAIT_3D_IDLE;
 
 	/* There are differences between R700 and EG in CP DMA,
 	 * but we only use the common bits here. */