passthrough sample destroys bound vertex buffers prematurely

[bjoeris]

Based on a Gapid trace of the passthrough sample, it looks like the vertex buffers are destroyed immediately after recording the command buffer, before submitting to the queue. They should not be destroyed until after all uses of the buffer have completed. Because it looks like these command buffers are submitted repeatedly, these buffers should not be destroyed at all.

This causes the passthrough sample to replay incorrectly with the new dependency graph-based dead code elimination going into Gapid.