[d3d12] Back-end Optimizations

[slomp]

• Finalize the implementation of the allocation cache (preferably through ID3D12Resource sub-allocations instead of just re-purposing d3d12_buffer objects in a free list)
• Favor ID3D12Device::CopyDescriptors() over re-creating them fresh with ID3D12Device::Create*View()
• On kernel dispatches, batch ResourceBarrier() calls in compute_barrier() instead of issue them one-by-one for each buffer
• Determine whether argument buffers are read-only, write-only, or read-write, in order to bind them to SRV slots instead of UAV; this should also allow for less wait_until_ when synchronizing buffers, and more intelligent resource transitions and barriers.