Support of GPU based instances in OpenStack

Summary
=======
Support of GPU based instances in OpenStack

Motivation
========
The usage of GPU is heavily increased in areas of audio transcoding for NFV, Video Processing and heavy computations for machine learning, there is a need of enabling GPU based instance creation. This is already supported in Public clouds like AWS and GCE and openstack needs parity for those interested in enabling this in Private clouds.

Description
=========
Currently, GPU is only supported in pass-through mode by libvirt + qemu and virtualizing GPU cards on the host is still not available. But we should have new nova scheduler filters so that nova can schedule the GPU flavors to the correct compute hosts. Also, the nova-compute should instruct the libvirt properly while the instance creation happens about the GPU required for the VM. We should also have new parameters in the nova flavor to support GPU based configuration.

Proposed Change
===============

All Nova components need to be checked to support this feature.

UX
===
There should be support on the Horizon to create GPU based instances.

Testing
======
To be planned.

Outside Dependencies
==================

To be analyzed.

Requirements Update Required
========================
Yes, we need to update the requirements one more formal discussion are started on this.

Doc Impact
=========
All documents needs to be updated with the new support.