fix: gpu memory leaks

new_buffer_design.txt

@@ -0,0 +1,9 @@
+The Post-Decoded Buffer should just be the ping pong ring buffer
+let's get some relationship in order
+- ping pong ring is per model
+- many camera may use the same model
+- this buffer is filled when we memcpy it from decode buffer
+but I need some more ground rules
+- in the model buffer, one frame per camera may be in the buffer, if older one of the same camera exist, evict it. This is a real time system so buffer should be as fresh as possible.
+- The goal of batching is not to pool up processing for the same camera but to pool up multiple camera.
+- if all camera in the pool already post its frame, flush the buffer too