See #872

This does away with the separate retained-reference command buffer
for the present call, and instead uses the regular 'deferred
release queue' for the default pass MTLRenderPassDescriptor.
Since the render-pass-descriptor holds a strong-ref to
swapchain surfaces this should make sure that those surfaces
outlive their command buffer when they are released by
MTKView because of a window resize. This means there's a memory
spike during window resize, but I guess there's no way around that.