Add support for serverless deployments

[RobertLucian]

Description

The cluster.yaml config would no longer specify an instance type for its deployments.

The instance type would be determined at deploy time, based on the amount of GPUs/CPUs/Mem/Infs/etc requested and based on the value of a field like instance_selection_policy: cost/performance. There would still be the option of overriding the instance type using the instance_type field in the cortex.yaml config.

Motivation

Eliminates the users' need to think of what resources they want and how they want it.

Notes

• This is related to Support concurrent CPU and GPU autoscaling groups #470 and Add spot vs on-demand to API configuration #970

[sp-davidpichler]

I think for this to usable (at least for gpu deployments) users need to be able to specify the set of instances available to the cluster. This is especially important if I have a model compiled with TensorRT and need a specific GPU arch.