Allow for multiple Inferentia ASICs to be assigned to each API replica

[RobertLucian]

Description

Remove the current limitation of 1 Inferentia ASIC per API replica. We're currently forced to go with only one because of some issue in the Neuron RTD.

Motivation

It will allow to partition models across multiple Inferentia ASICs.

Additional context

As reported in aws-neuron/aws-neuron-sdk#110.

[RobertLucian]

This is now possible thanks to https://github.com/aws/aws-neuron-sdk/blob/master/docs/neuron-container-tools/k8s-neuron-scheduler.md.