Add device map config

[mrwyattii]

In MII-Legacy we had deploy_rank which would allow us to specify which GPUs to deploy a model to. This did not compose well with multiple replicas, so we I've refactored that code and brought it into the latest MII.

Here we had a device_map to the config that allows users to specify which devices they want to deploy a model to for the persistent deployment (mii.serve). This works with multi-replica and multi-node cases. We can provide the following types to device_map:

• int: device_map = 1 - deploy single GPU model to GPU1
• List[int]: device_map = [2,3] - deploy a 2 GPU model to GPU2 and GPU3
• List[List[int]]: device_map = [[0,2],[1,3]] - deploy 2 dual-GPU replicas, one to GPU0 & GPU2, the other to GPU1 & GPU3
• Dict[str,List[List[int]]]: device_map = {"host0":[[0,1],[2,3]], "host1":[[0,1],[2,3]]} - deploy 4 dual-GPU replicas across 2 nodes

The default value is "auto" which will automatically place models / replicas across devices and nodes. Users must still specify the proper replica_num and tensor_parallel values, and these values must match with the device map provided. The device map is not required and is only needed when the non-default model/replica placement is not desired.

resolves #283